Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
Improving Skin Cancer (Melanoma) Detection: New Method
By
Mohamed Khaled Abu Mahmoud
Abstract
1
Abstract
Melanoma, the deadliest form of skin cancer, must be diagnosed early for effective
treatment. Rough pigment network and qualities are important signs for melanoma
diagnosis using pathologist images. The main focus of this thesis is to improve skin cancer
(Melanoma) detection through introducing novel image processing approach for a
computer-aided system based on pigment network and elements detection on pathology
images. It is important to propose an automated system for differentiating between
melanocytic nevi and malignant melanoma. This thesis describes a novel image processing
approach for computer-aided pigment network and elements detection on dermoscopy /
pathology images. The proposed methods provide meaningful ideas of structures, and
extract features for melanoma detection. Additionally, the thesis presents efforts towards
prevention of melanoma, by developing a smart system to locate pigment networks.
The thesis aims to cover a complete theoretical model for simulating the processes that
takes place when a human interprets an image generated by the eye, through designing a
reliable system, that can provide a screening method that “filters” lesions and melanoma
in a general practice. The proposed system is to be used with a standard PC with input
from a high quality digital camera, dermoscopy / microscopy slides or any other suitable
hardware sources. This system analyses the structure of a mole / skin defects, detects
cancer, identifies features, makes a decision and provides the result.
The result of the proposed system shows that the Skin Cancer (Melanoma) Detection
strategy which uses SVM performs reasonably satisfactorily (accuracy 77.44%, sensitivity
83.60 %, and specify 70.67%). Furthermore, the SVM based wavelet Gabor (SVM-WLG)
performs better than the SVM (81.61%, 88.48%, and 74.51 % accuracy, sensitivity, and
specify respectively). However, the Swarm-based SVM (SSVM) performs better than the
other two algorithms, with average for accuracy, sensitivity, specificity of 87.13%, 94.1%
and 80.22%, respectively.
Originality
2
Certificate of Authorship / Originality
I certify that the work in this thesis has not previously been submitted for a degree nor has
it been submitted as part of requirements for a degree except as fully acknowledged within
the text.
I also certify that the thesis has been written by me. Any help that I have received in my
research work and the preparation of the thesis itself has been acknowledged. In addition, I
certify that all information sources and literature used are indicated in the thesis.
Mohamed Khaled Abu Mahmoud
November, 2014
Acknowledgments
3
Acknowledgments
4
Acknowledgments
Firstly, I would like to express my sincere gratitude to my principal supervisor, Assoc.
Prof. Adel Al-Jumaily, who provided such expert guidance and advice throughout this PhD
candidature. He possesses an amazing amount of expertise and an ability to critically
appraise work, that ensures the end result is worthy and importantly of value to public
health.
I would also like to sincerely thank and acknowledge the pathologist Dr. Sarbar Napaki
from Southern Pathology laboratory for providing us with pathologist data samples and for
updating our medical expertise with the intension of improving our results. Also I would
like to extend my thanks to Dr. Ryad Ahmed (for reviewing Chapter 2; Human Skin
Biology and Cancer and Related Work in this thesis) and Dr. Rami Khushaba, Dr. Jebrin
Sharawneh and Miss. Hayat Al-Dmour who have supported me whenever requested.
I am truly blessed to work in research for an organization dedicated to cancer detection and
due to this I would like to express my sincere gratitude to University of Technology for the
great experience and knowledge it gave me, my colleagues at the University of
Technology, Sydney, Faculty of Engineering and Information Technology, School of
Electrical, Mechanical and Mechatronic Systems who always encouraged and supported
me, providing that little extra push to enable me to complete this thesis.
I am very lucky to have a wonderful supportive family. My two sons, my three daughters,
my three sons in law and my grandkids have always encouraged me in my striving to
complete this piece of work. Finally, to my wife, my partner in life for the past 43 years,
this would not have been possible without your love and support.
Table of Contents
5
Table of Contents
Table of Contents
6
Table of Contents
7
Extended Table of Contents
Table of Contents
Table of Contents
8
Table of Contents
9
Table of Contents
10
Table of Contents
11
List of Figures
12
List of Figures
Figure 2.1 shows, Cells structure; source: www.getting-in.com ........................................ 33
Figure 2.2 The eukaryotic cell cycle, G Phases: growing, S phase: synthesis/ DNA replication, G2 Phase: Growing and preparing for Mitosis, M mitosis. .............................. 34
Figure 2.3 The cell cycle; G0 terminally differentiated cell G1 gap phase 1 G2 gap phase 2 M mitotic phase S synthesis phase ...................................................................................... 35
Source: Wheatear’s Functional Histology: A Text and Colour Atlas, 5th Ed ..................... 35
Figure 2.4 H&E-stained section of skin. High magnification (40xs) view. Haematoxylin stains the nuclei of cells blue to bluish-purple, and eosin stains other cellular elements in the tissues from pink to red [43]. ......................................................................................... 35
Figure 2.5 Displayed Cross Section & the main layers of Human Skin ............................. 37
Source: nurrashidah2204.blogspot.com ............................................................................... 37
Figure 2.6. Skin Color Distribution around the World ....................................................... 38
Source: www.gbhealthwatch.com ....................................................................................... 38
Figure 2.7. a) Cross section of human skin, b) Structure of Epidermis. ............................. 39
Figure 2.8. The Layers of Skin - the Epidermis and Dermis (At the top, the close up shows a squamous cell, basal cell, and melanocyte). ..................................................................... 42
Figure2.9- a) Basal Cell Carcinoma (BCC), b) Squamous Cell Carcinoma (SCC) ........... 42
Figure 2.10 – Melanoma ..................................................................................................... 43
Figure 2.11: Malignant Melanoma, World Age-Standardised Incidence Rates, World Regions, 2008 Estimates. ..................................................................................................... 44
Figure 2.12. Lentigomaligna Melanoma. ............................................................................ 45
Source: melanomaknowmore.com ...................................................................................... 45
Figure 2.13. Superficial Spreading Melanoma. .................................................................. 46
Source: melanomaknowmore.com ...................................................................................... 46
Figure 2.14. Nodular Melanoma. ........................................................................................ 46
Source: melanomaknowmore.com ...................................................................................... 46
Figure 2.15.Stages of Cancer Development and Metastasis. .............................................. 47
Source: http://www.cancervic.org.au/about-cancer/advanced-cancer ................................. 47
Figure 2.16: Five-year relative survival from selected cancers by remoteness area, Australia. 2006-2010. ......................................................................................................... 49
Figure 3.1 - Show the function of the eye. .......................................................................... 55
Source: http://cdn2.hubspot.net/hub/60407/file-350482438-jpg/images/Master_Eye_logo_short_tag_line_revised_10-13-2013-resized-600.jpg ......... 55
Figure 3.2 - Distance of vision. ........................................................................................... 55
List of Figures
13
Source:hyperphysics.phy-astr.gsu.edu/hbase/vision/accom.html ........................................ 55
Figure 3.3 - Structure of the Eye. ........................................................................................ 55
Source:drpion.be/en/bouw-van-het-oog.htm ....................................................................... 55
Figure 3.4 Focal length defined [When parallel rays of light strike a lens focused at infinity, they converge to a point called the focal point. The focal length of the lens is then defined as the distance from the middle of the lens to its focal point.] ............................... 59
Figure 3.5. Displayed, similarities of overall design in the principal features of an optical microscope, a transmission electron microscope and a scanning electron microscope. ...... 63
Source: www.nslc.wustl.edu ................................................................................................ 63
Figure 3.6. Principle of CSLM. Note the same optical path is used for the detector and the source. Optics are used to direct the light towards the detector .......................................... 64
Figure 3.7. Two-step procedure for the classification of pigmented skin lesions. Adapted from Argenziano.[122] ........................................................................................................ 65
Figure 3.8. Algorithm for the determination of melanocytic versus non melanocytic lesions according to the proposition of the Board of the Consensus Netmeeting. Adapted from Argenziano.[122] ........................................................................................................ 66
Figure 3.9. A, Macroscopic picture of a superficial spreading malignant melanoma (Breslow thickness 0.52 mm; Clark level II). B, Dermoscopy of A shows (atypical) pigment network and branched streaks and can therefore be considered a melanocytic lesion. ................................................................................................................................... 66
Figure 3.10. A, Macroscopic picture of a blue nevus. B, Dermoscopy of A shows steel-blue areas (no pigment network, no aggregated globules, and no branched streaks). ......... 67
Figure 3.11. A, Macroscopic picture of a seborrheic keratosis. B, Dermoscopy of A shows comedolike openings (a), multiple milia-like cysts (b), and fissures (c). ............................ 67
Figure 3.12. A, Macroscopic picture of a seborrheic keratosis. B, Dermoscopy of A shows comedolike openings and multiple milia-like cysts. ............................................................ 67
Figure 3.13. A, Macroscopic picture of a basal cell carcinoma. B, Dermoscopy of A shows maple lifelike areas, ovoid nests, and arborized telangiectasia. .......................................... 67
Figure 3.14. A, Macroscopic picture of a basal cell carcinoma. B, Dermoscopy of A shows multiple spoke wheel areas. ................................................................................................. 67
Figure 3.15. A, Macroscopic picture of an angioma. B, Dermoscopy of A shows red lagoons. ................................................................................................................................ 68
Figure 3.16. Asymmetry Border Color Diameter Evolution (ABCD-E) rule for Diagnosis of Melanoma ........................................................................................................................ 68
Source: www.webmd.com ................................................................................................... 68
Figure 3.17 - Normal moles and Melanomas [Benign and Malignant]. ............................. 70
Source: www.skinbychar.com ............................................................................................. 70
Figure 3.18: Showed: a) Melanoma in skin biopsy with H&E Stain (“This case may represent superficial spreading melanoma.”), b) Lymph node with almost complete replacement by metastatic melanoma. The brown pigment is focal deposition of melanin. ............................................................................................................................................. 71
List of Figures
14
Figure 3.19. Diagnosis of Melanoma using Three-point Checklist .................................... 72
Figure 3.20 - Example of pigmented skin lesion. Left: Traditional imaging technique. Right: Dermoscopy imaging technique. .............................................................................. 73
Figure 3.21. (a) Macroscopic Image of a Lesion (b) Dermatoscopic Image of the same Lesion. Source: www.jle.com ............................................................................................. 74
Figure 3.22- Show the results of skin lesion boundary tracing algorithm, my data experiment shown [a] Image center of mass, [b] Image process ......................................... 75
Figure 3.23- From left to right: the original image (source image from: http://ijplugins.sourceforge.net/plugins/clustering/), clustered image using 2 clusters (poor), clustered image using 3 clusters (close but one key cluster is missing), clustered image using 4 clusters (optimal), and clustered image using 10 clusters (artifacts are noticeable). ........................................................................................................................... 77
Figure 3.24- Structure of fuzzy image processing .............................................................. 78
Figure 3.25- Steps of fuzzy image processing. ................................................................... 78
Figure 3.26. Manual Segmentation by trained pathologists using the ‘Aperio Image Scope’ software [156]. ..................................................................................................................... 82
Figure 3.27 Showed the Radial growth phase melanoma. .................................................. 83
Figure 3.28 showed the Vertical growth phase melanoma ................................................. 84
Figure 3.29 Microscopic satellites, (a). Shows neoplastic group is discontinuous (arrow) from the overlying vertical growth phase component. Shows the melanocytes must have a malignant cytomorphology (b). ........................................................................................... 85
Figure 3.30 Regression: Regression of over 75% of the tumor volume of a melanoma is considered a bad prognostic sign. ........................................................................................ 86
Figure 3.31: a) An image of a lesion under clinical view (naked eye). b) Shows the same lesion under a dermoscopy with oil immersion. .................................................................. 86
Figure 3.32: a) Colors allow the physician, to some extent, to draw conclusions about the localization of pigmented cells within the skin. Black and brown indicate pigmentation in the epidermis, while grey and blue correspond to pigmented cells within the superficial and deep dermis, respectively [185]. .......................................................................................... 87
Figure 3.33: Figures (a, b, c, d, and e) show analogue dermoscopy. All of them, except (a), are attachable to digital cameras to function as digital dermoscopy. Figures (f, g, and h) show DinoLite, Handy scope, and Dermoscopy which are modern digital dermoscopy. ... 88
Figure 3.34. The Solar Scan instrument: (a) global appearance, (b) camera, (c) user interface [192]. Source: www.medgadget.com ................................................................... 90
Figure 4.1 Showed: Four stages CAD system for skin lesions. .......................................... 96
Figure 4.2 true color Pathological Digital Image. ............................................................... 96
Figure 4.4, calculating the median value of a pixel neighborhood. .................................. 102
Figure 4.5. The basic structure of the MACWM filter ..................................................... 104
Figure 4.6. Shows the output image of Adaptive median Filtering. ................................. 104
Figure 4.7 Showed: (a) grayscale image, (b) Histogram out. ........................................... 107
Figure 4.8. Showed: intensity image using ....................................................................... 109
List of Figures
15
Figure 4.9. Sobel Edge Detection Image .......................................................................... 114
Figure 4.10.Binary Output Image Otsu’s Method[ ........................................................... 114
Figure 4.11, selected thresholds in valleys between peaks ............................................... 116
Figure 4.12, obtaining the best possible separation measure for two classes ................... 116
Figure 4.13, selecting a threshold by visually analyzing a bimodal histogram. ............... 116
(Principle of histogram peak separation). .......................................................................... 116
Figure 5.1 displayed the main categories: texture elements, regularity, randomness, directionality and regularity. .............................................................................................. 126
Figure 5.2, showed, (a) images (5x 5) matrix with 3 grey levels 0, 1, and 2. (b) The co-occurrence matrix for d= (1, 1). ......................................................................................... 127
Figure 5.3, displayed the response from convolving image sample with filter. ............... 129
Figure 5.4, (a) Wavelet decomposition of an image (b) Block diagram of the decomposition of an image ................................................................................................ 130
Figure 5.5 displayed a block diagram of the discrete wavelet transforms (DWT – Gabor approach). .......................................................................................................................... 130
Figure 5.6, the three principal approaches of feature Mammographic e selection. The shades show the components used by the three approaches: filters, wrappers and embedded methods. ............................................................................................................................. 133
Figure 5.7: Two-out-of-many separating lines; (a) with smaller margin and (b) with larger margin ................................................................................................................................ 138
Figure 5.8: Margin m between two supporting hyperplanes. ........................................... 139
Figure 5.9: Illustration of mapping using a transform . ............................... 140
Figure 5.10: Introducing slack variable in soft-margin SVM ........................................ 142
Figure 5.11: Loss Functions .............................................................................................. 144
Figure 5.12: Melanoma detection using swarm based support vector machine. .............. 145
Figure 5.13: The pseudo of the PSO for the SVM parameter optimization ..................... 146
Figure 5.14: The particles of the PSO ............................................................................... 147
Figure. 6.2 Skin Cancer Image, (a) Original RGB true colour image .............................. 153
Figure 6.3. The basic structure of the MACWM filter ..................................................... 154
Figure 6.4. A snake with traditional potential forces cannot move into the concave boundary region. ................................................................................................................ 156
Figure 6.5 A snake with GVF external forces moves into the concave boundary region. 157
Figure. 6.6. Skin Cancer greyscale image showing: (a) Wiener2 filter has removed the spots effectively (b) noisy image filtered by the median filter .......................................... 157
(a) Without filtering, (b) with filtering. ......................................................................... 158
Figure.6.7 Skin Cancer Image Histogram, ........................................................................ 158
Figure. 6.8 Skin Cancer greyscale image showing: (a) Sobel segmented image ............. 159
(b) Gradient vector flow (GVF) segmented image ............................................................ 159
List of Figures
16
Root mean square error ...................................................................................................... 160
Figure 6.9. Result of segmentation by threshold with GHE (bottom right) and LHE (bottom left), Top: shows histogram Local and Global results ......................................... 162
Figure 6.10.Segmentation by Thresholding (ROI) ........................................................... 162
Figure 6.11. Segmentation by SRM .................................................................................. 163
Figure 6.12. Display of the image and its transform (wavelet coefficients) ..................... 164
Figure. 6.13: A single curvelet with width and length . .............................. 164
Figure 6.14. Rectangular frequency (basic digital) tiling of an image with 5 level curvelet. ........................................................................................................................................... 165
Figure 6.15. Curvelet demising image that contains oriented texture and cartoon edges (a) Original (b) Noisy, and (c) Curvelet. ................................................................................. 166
Figure. 6.16. The results showed, original image, image after median filter, gray scale image, ................................................................................................................................. 172
Figure. 6.17. - A Wavelet Packet decomposition tree ...................................................... 178
Figure. 6.18 shows the melanoma detection employing SVM (SVMR, SVMP, and SVML); the input is pathology images and the output is melanoma (-1) or benign (+1). 181
Figure A1, showed a physician's hands are seen performing a needle biopsy to determine nature of lump either fluid-filled cyst or solid tumour ...................................................... 189
[This image was released by the National Cancer Institute, an agency part of the National Institutes of Health, with the ID 1973 (image)] ................................................................. 189
Figure A2, the diagnostics showed: Micrograph of a needle aspiration biopsy specimen of a salivary gland showing adenoid cystic carcinoma. (Source: Pap stain. MeSH D044963). ........................................................................................................................................... 189
Figure A3, showed a Normal Epidermis and Dermis with Intradermal Nevus 10x-cropped. (Source, Kilbad). ................................................................................................................ 190
Figure A4, this image shows a crossection of the vascular tissue in a plant stem. Showed as an example for bright field micrograph. (Source: Wikipedia). ..................................... 190
Figure A5, Principle of confocal microscopy ................................................................... 191
Figure A6 showed the nucleolus is contained within the cell nucleus. (Source: Wikimedia Foundation, Inc.) ................................................................................................................ 192
Figure A7 displayed the negative of the logarithm (base 10) of the optical Density (OD). (Source: [email protected]) ................................................................................... 193
Figure A8, showing a pathologist examines a tissue section for evidence of cancerous cells while a surgeon observes. (Source: wikipedia.org) ........................................................... 193
Figure A9, showed: a) A TEM image of the polio virus is 30 nm in size, b) Layout of Optical components in a basic TEM. (Source: Wikimedia)Appendix B: Optical Density of Transmission Microscopy Images ..................................................................................... 194
Figure B1. Example optical density image ....................................................................... 196
Figure C1, showed a graph with three connected components ....................................... 197
List of Figures
17
Figure C2, displayed the machine learning and data mining which used to solve problem in the areas of Classification, Clustering, Regression, Anomaly detection, Association rules, Reinforcement learning, Structured prediction, Feature learning, Online learning, Semi-supervised learning, and Grammar induction. ......................................................... 198
Figure D1: Margin m between two supporting hyperplanes ............................................. 203
List of Tables Introduction
18
List of Tables
Table 2.1: Breslow's depth .................................................................................................. 48
Table 2.2: Survival figures from British Association of Dermatologist Guidelines 2002 .. 48
Table 2.3: Observed incidence (2009), and morality (2010) of melanoma of the skin and estimated for 2012. .............................................................................................................. 51
Table 2.4: Observed incidence (2009), and morality (2010) of non-melanoma of the skin and estimated for 2012 ......................................................................................................... 52
Table 6.1 Sensitivity and Specificity comparison. ............................................................ 152
Table 6.2 Comparison between proposed operators (Sobel, Roberts, Prewitt, Laplacian, Canny and Otsu) and Gradient vector flow (GVF) segmented images. ............................ 160
Table 6.3 BNN Classification Test with Different Wavelet. ............................................ 161
Table 6.4. BNN Classification Test for SRM & Thresholding. ........................................ 163
Table 6.6. Displaying the comparison between wavelet and curvelet features results. .... 166
Table 6.7. svm classification for all the dataset (including dermoscopy images) with features extracted from wavelets ....................................................................................... 169
Table 6.8. svm classification excluding dermoscopy images with features extracted from wavelets ............................................................................................................................. 170
Table 6.9. svm classification for all the dataset (including dermoscopy images) with features extracted from curvelets ....................................................................................... 170
Table 6.10. svm classification excluding dermoscopy images with features extracted from curvelets ............................................................................................................................. 170
Table 6.11. sensitivity and specificity statistics [310] ...................................................... 171
Table 6.12. Described sensitivity and specificity .............................................................. 174
Table 6.13. show result after training or the (svm) network. using (sfs) technique ......... 174
Table 6.14. the result after training of the (svm) network with using (sfs) technique. ..... 175
Table 6.15. the result after training of the (svm) network without using (sfs) technique. 175
Table 6.17 The Result after Training of the (SVM) Network, with using (SFS)Technique. ........................................................................................................................................... 180
Table 6.18 The Result after Training of the (SVM) Network. with using (SVM+WLG+PSO) ........................................................................................................... 180
Table 6.19 Described Sensitivity and Specificity [31] .................................................... 181
Introduction
19
CHAPTER 1
Introduction
Australia has one of the highest rates of skin cancer in the world, at nearly four times the
rates in Canada, the US and the UK [1]. Two in three Australians will be diagnosed with
skin cancer by the time they are 70. It has been estimated 115,000 new cases of cancer per
year are diagnosed and more than 43,000 people are expected to die from cancer per year
according to the Cancer council of Australian 2014 [2]. The Chair of Public Health
Committee pronounced that more than 430,000 cases are treated for non-melanoma, and
more than 10,300 people are treated for melanoma, with 1430 people dying each year[2].
Skin cancer is the most expensive cancer [3]. In 2001, it was estimated the treatment of
non-melanoma skin cancer cost $264 million and melanoma $30 million. The skin cancer
malignant melanoma is the deadliest form and accounts for 75% of all cancer deaths [4]
[5]. It can be removed by simple surgery if it has not entered the blood stream. The annual
rate of melanoma is increasing at the rate of 6% Worldwide[5]. If the melanoma gets
deeper than 3 millimeters[6], the chance of survival is 59%[7]. A patient can recover from
melanoma if it is detected and treated in the early stages and this can achieve cure ratios of
over 95%. Early diagnosis is obviously dependent upon patient attention and accurate
assessment by a medical practitioner. The variations of diagnosis are sufficiency large and
there is a lack of detail of the test methods. The diagnostic is mostly based on visual
inspection. The traditional imaging is just a recording of what the human eye can see using
a digital camera while the Dermoscopy known as Epiluminescence Light Microscopy
(ELM) images need a professional with experience to get the required image.
Australia is one of many countries in which skin cancer is widely spread in comparison to
other types of cancer[8], and researchers from the School of Medicine, University of
Queensland, Australia, found that Australian melanoma rates are the highest globally at
almost four times the rates seen in Canada, the United Kingdom and the United States[9]
[10] with mortality two-fold that of Southern and central Europe[11].
CHAPTER 1 Introduction
20
1.1. Research Aims
The aims of this research are to improve the quality of existing diagnostic systems by
proposing 283 new feature extraction, selection and classification methods. These methods
are aimed for early detection of pathologic melanoma. This process provides a better tool
for screening and detecting lesions that are considered to be suspicious allowing early
treatment and improved survival rate. The research focuses on the proposed algorithm that
is a combination of segmentation methods for skin cancer image to detect the lesion border
(edges), followed by feature extraction, selection and classification. The diagnostic results
of the proposed system are compared with other algorithms as well.
1.2. Problem Description
It is difficult to visually differentiate a normal mole from an abnormal one and general
practitioners (GP) do not usually have sufficient expertise to diagnose skin cancers. Skin
cancer specialists will significantly improve the identification rate by over 75% but are
often severely overloaded by referrals from regional practices. For this reason the thesis
focuses on the development and implementation of a skin cancer screening system that can
be used in a general practice by non-experts to ‘filter’ the cases, so that in the case of an
abnormal mole, a patient can be referred to a specialist.
The automatic diagnostic system for melanoma has been developed and designed for use
with a standard PC with input from a quality digital camera, / dermoscopy / microscopy
slides or any other suitable hardware sources. It analyses the structure of skin defects,
detects cancer identifying- features, makes a decision using a knowledge database and
outputs a result. Skin cancer experts create a knowledge database by training the system
using a number of case-study images.
One commercial product, Solar Scan by Polar technics, has an accuracy rate of 92%. Solar
Scan is a complex system, taking high quality Epiluminescence Light Microscopy (ELM)
images and using advanced image analysis techniques to extract a number of features for
classification which makes it not suitable for normal person usage. This Dermoscopy
equipment (known as Epiluminescence Light Microscopy) needs a professional with
experience to get the required image. Researchers in [12] [13] [14] reported that the
performance improved by using Oxford Recognition Skin Cancer Screen System
CHAPTER 1 Introduction
21
(ORSCSS), which was developed by Oxford Recognition Limited (ORL) in collaboration
with Loughborough University. This process investigates the applications of pattern
recognition using fractal geometry as a central processing kernel, which allowed the design
of a new library of pattern recognition algorithms including the computation of parameters.
Also this process of object detection, recognition and classification in a digital image
where the classification method is based on the application of a set of features which
include fractal parameters where it incorporates the characterisation of an object in terms
of its texture. However, reported that this texture based analysis alone is not sufficient in
order to design a recognition and classification system [12].
A further development for this process is to consider the effect of replacing the fuzzy logic
engine used with an appropriate Artificial Neural Network (ANN) even then it is not clear
whether the application of an ANN could provide a more effective system and whether it
will provide better flexibility according to the type of images used and the classifications
that may be required. However, the result of the proposed system shows that the Skin
Cancer (Melanoma) Detection strategy which uses SVM performs reasonably satisfactorily
(accuracy 77.44%, sensitivity 83.60 %, and specify 70.67%). Furthermore, the SVM based
wavelet Gabor (SVM-WLG) performs better than the SVM (81.61%, 88.48%, and 74.51 %
accuracy, sensitivity, and specify respectively). However, the Swarm-based SVM (SSVM)
performs better than the other two algorithms, with average for accuracy, sensitivity,
specificity of 87.13%, 94.1% and 80.22%, respectively. It was concluded that the proposed
system gives fast, accurate classification and better results. This thesis concludes that there
are some possible factors to improve the accuracy of detecting malignant melanoma by
having a higher number of images for training of the SVM network [15].
1.3. Objectives
This thesis targets the above mentioned problems in an attempt to solve them by
developing new algorithms and applying these algorithms into novel applications. This
thesis aim to develop hybrid artificial intelligence techniques employing two or more of
the following methods: Wavelet Gabor (WLG), fuzzy logic (FL), neural networks (NN),
k-means clustering, hybrid evolutionary optimization algorithm based on combining
Particle Swarm Optimization (PSO) and Support Vector Machine (SVM) called PSO-
SVM. The process involves:
CHAPTER 1 Introduction
22
1. Using combination of methods to target different parts of the automated recognition
system like segmentation, features extraction, feature selection, dimensionality
reduction techniques, and pattern recognizers or classifiers.
2. Comparing with the available approaches in literature and demonstrate the
successfulness and superiority of the new novel methods.
3. Investigating the utilization of the pattern recognition system into novel applications in
healthcare, like identifying the deadliest form of skin cancer problems.
4. Using real patient pathology images sourced from a consultant Pathologist.
This thesis will include a practical element to investigate questions, primarily targeted at
skin cancer (melanoma) patients. The thesis is trying to direct the work to meet the
Australian community needs in the area of skin cancer (melanoma).
1.4. Methodology
Most general practitioners (GP) have less experience in the full range of melanoma forms,
which means many cases, are not diagnosed properly [GPs cannot provide cancer care
alone. Nor does the solution lie in the recruitment of more specialists, such as
dermatologists [16]. Last decade researchers developed automatic early diagnostic aided
systems to advance the diagnostic accuracy from 60% to 92% [17] and they became able to
provide recommendations for non-specialized users. As mentioned above early detection
improves survival rate. The methodology that has been developed relies on extracting and
selecting specific information features that can be used to distinguish malignant and benign
lesions by setting an automated cancer diagnosis / image processing system. The common
approach to skin lesion images combines four main computational intelligent stages [18] as
combined on the following modules:
Pre-processing stage: This stage consists of filtering and contrast enhancing techniques to
remove any unwanted structures that might corrupt the image. Also the aim of this stage is
to eliminate the background noise and improve the image quality for the purpose of
determining the focal areas in the image.
CHAPTER 1 Introduction
23
Segmentation using thresholding, region based approach or boundary tracing algorithms to
localize the lesion.
Feature extraction / selection. This stage quantifies the distribution of the cells across the
tissue. It primarily makes use of the grey-level dependency of the pixels and selects the
useful information from these images and passes it to the classification stage for training
and testing in order to distinguish between malignant or benign lesions.
The classification of image: This stage used Back-propagation Natural Network or Support
Vector Machine [19] to find out which algorithm performed better. This step uses
statistical analysis of the features and machine learning algorithms to reach a decision.
The variations of diagnosis are sufficiently large and there is a lack of detail of the test
methods. This thesis looks at developing a practical approach to extend the functionality of
the existing methods and tools for medical diagnostic purposes and in the long run provide
an alternative basis for researchers to experiment with new and existing methodologies for
skin cancer detection and diagnosis.
1.5. Contribution of the Doctoral Research
The purpose of this research is to improve the quality of existing diagnostic systems by
proposing some new feature extraction, selection and classification methods. These
methods are aimed for early detection of pathologic melanoma. This research focuses on
the proposed algorithm that is a combination of segmentation methods for skin cancer
image to detect the lesion border (edges), followed by feature extraction, selection and
classification. The methodology that has been developed distinguishes malignant and
benign lesions by setting an automated cancer diagnosis / image processing system.
This thesis, presents a segmentation of skin cancer images based on gradient vector flow
(GVF) snake algorithm. This algorithm used winner and median filters respectively to
remove the unwanted noise and hairs from the image. Since edge detection is the initial
step in object recognition, it is necessary to know the differences between edge detection
algorithms. In this thesis we classified the most commonly used algorithms into five
categories, then these algorithms have been applied to 8 images. A comparative way
published in our paper in [20] that shows, Sobel and Prewitt with better performances for
an image contaminated with Gaussian noise and filtered with Winner and Median filters.
CHAPTER 1 Introduction
24
These operators also have good contrast, edge map strength and low noise content. Prewitt
is more acceptable than Roberts and Laplacian while Sobel are more acceptable than
Prewitt. It also shows that Laplacian is very much sensitive to noise. Canny operator shows
better detection, good response and improving signal to noise ratio. Filtering the noisy
image with a suitable filter is an initial process in the edge detection for noisy images. The
subjective evaluation of edge detected images show that proposed operators , have
evaluated both the performance and Root mean Square of Signal to Noise ratio between the
input image (operators) and the output image (GVF).
Building an image classification system for early skin cancer detection includes different
stages. The pre-processing resizes the image which improves the speed performance and
removes the extra features. Post-processing enhances the image quality and sharpens the
outline of the cancer cell. On other hand, segmentation to find the region of interest by
removing the healthy skins from the image and keeping the cancer cell whereas the feature
extraction decomposes the useful features without prior clinical knowledge.
The classification results, which based on a database of mixed type of images, showed in
our publication [21] are as follows: 58.44 % for back-propagation neural network when
using Wavelets and 86.57 % when using Curvelet. This thesis concludes that there are
some possible factors for low classification results; one of the main factors that affect
reaching high accuracy is the type of database used. A mix of images that were acquired
from different sources mainly normal digital images and dermoscopy images has been
used. While the proposed technique has tried to generalize the system working area, it has
traded off the accuracy. A possible improvement and generalization can be achieved by
dealing with the effect of large variation and overcoming it using larger databases.
In this thesis, the Support Vector Machine (SVM) has been implemented for classification
of benign from malignant skin tumor. MATLAB package is used to implement the
software in the current work; these features were carried out to generate training and
testing of the proposed SVM. The present work is a new application based on histo-
pathological images of skin lesions that required finding out new features and the
correlation of the reduced numbers of features and getting better accuracy. It was required
to modify many of the mentioned techniques to make them working for such application.
42 images (28 benign images and 14 melanoma images) were sampled and split into a
CHAPTER 1 Introduction
25
training set (mixed 28 images) and test set (mixed 14 images) were sampled from
microscopic slides of skin biopsy which have been used in the current work. The
performance of the algorithm was evaluated by computing the percentages of Sensitivity
(SE) equation (1.1), Specificity (SP) equation (1.2) and Accuracy (AC) equation (1.3): the
respective definitions are as follows:
1.1
1.2
1.3
Where TP is the number of true positives (expects malignant as malignant), TN is the
number of true negatives (expects benign as benign), FN is the number of false negatives
(expects malignant as benign), and FP is the number of false positives (predicts benign as
malignant). To estimate the performance of classifier based on the classification of benign
and malignant skin cell nuclei, sensitivity, specificity and accuracy of prediction have been
calculated according to the above equations for all of the testing data.
The higher accuracy of diagnosis proposed work is calculated to display. The results
proved to be one of the best results comparing with expertise / physicians results displayed
in our publication in [22]. The obtained accuracy of the system is 88.9 % whereas the
sensitivity and specificity were found to be equal 87.5 % and 100% respectively. By
comparing the results in [22], it was concluded that the proposed system gives fast and
accurate classification and better results of skin tumors than physicians do. This thesis
concludes that there are some possible factors to improve the accuracy of detecting
malignant melanoma by having a higher number of images for training of the SVM
network.
A Novel feature extraction methodology based on histopathalogical images and subsequent
classification by Support Vector Machine was used. The results show that using the
sequential forward selection (SFS) technique produces one of the best results compared
with expertise / physicians results displayed in our paper in [23]. The obtained accuracy of
the system is 81 % whereas the sensitivity and specificity were found to equal 76 % and
100% respectively. through comparing the results in [23] and from our experimental
results CPU time which is faster than the previous experiment it produced more accurate
classification results, than physicians do.
CHAPTER 1 Introduction
26
The contribution of this thesis proposes an automated non-invasive system for skin cancer
(melanoma) detection based on Support Vector Machine classification. The proposed
system uses a number of features extracted from the Wavelet or the Curvelet
decomposition of the grayscale skin lesion images and colour features obtained from the
original colour images. The recognition accuracy obtained by the Support Vector Machine
classifier used in this experiment is 87.7 % for the Wavelet based features and 83.6 % for
the Curvelet based ones. The proposed system has resulted in a sensitivity of 86.4 % for
the case of Wavelet and 76.9% for the case of Curvelet. It has also resulted in a specificity
of 88.1% for the case of Wavelet and 85.4% for the case of Curvelet. The obtained
sensitivity and specificity results are comparable to those obtained by dermatologists[24].
A hybrid system for skin lesion detection: based on Gabor wavelet and support vector
machine (SVM) for classification of benign from malignant skin tumor has been
implemented in this thesis. 79 images sampled from microscopic slides of skin biopsy have
been used in the current work which uses a new application based on histo-pathological
images of skin lesions that required finding out new features getting the reduced numbers
of features and getting better accuracy. It was required to modify many of the mentioned
techniques to make them work for such an application. The higher accuracy of diagnosis of
the proposed work is calculated and displayed. The results show that using the particle
Swarm Optimization Support Victor Machine (SSVM) to improve the parameters of
support victor machine (SVM) technique produces one of the best results compared with
expertise / physicians results displayed in [25]. The obtained accuracy of the system is
87.1% whereas the sensitivity and specificity were found to equal 94.1% and 80.2%
respectively. The proposed system produced more accurate classification results than
physicians did [25].
This doctoral research has provided contributions to Skin Cancer (Melanoma) Detection
models. Novel algorithms and SSVM, a particle swarm optimization (PSO) was used to
optimize the SVM parameters that have been introduced for Melanoma detection.
Considering the results provided by this doctoral research, several peer reviewed papers
have been published, either in journal paper or conference papers. The full–reviewed
conference papers are presented in conferences organized by IEEE, Engineering in
Medicine & Biology and IEEE Computational intelligence Society’s
CHAPTER 1 Introduction
27
Through this thesis it was discovered that there are some possible factors to improve the
accuracy, feature extraction and sequential feature selection to choose the most relevant
features to get higher skin cancer detection accuracy of detecting malignant melanoma by
having a higher number of images for training of the SVM network. Future research aims
to develop hybrid artificial intelligence techniques employing the following genetic
algorithms; hybrid of fuzzy inference system (FIS), Ant Colony Optimization (ACO),
PSO-ACO, Ontology, Contourlet Transform and others in multifunction pattern
recognition systems
1.6. Structure and Summary of the Dissertation
This dissertation consists of 6 chapters including discussion and conclusions, appendixes
and references. The contents are organized sequentially from the first chapter until the last
chapter. The contents commence with the introduction of melanoma and finish with
discussion and conclusion.
Chapter 2: This chapter Human Skin Biology and Cancer presents and discuss the structure
and functions of the human tissues and cells. It takes into account the understanding of the
study of histology and how microscopes function. It also includes the various staining
procedures used to prepare tissues to make them visible through the microscope. Also
discussed is the histology, the study of the microscopic anatomy of cells and tissues of
humans, its role as an essential tool of biology and medicine. Histology is commonly
performed by examining cells and tissues through sectioning, staining and examination
under a light microscope or electron microscope. The study of changes in cell anatomy is
known as histopathology. A sound knowledge of the normal structure of skin is essential
for an understanding of pathology.
Chapter 3: This chapter presents an Image Processing and Automatic Skin Cancer
Diagnosing relevant to melanoma detection. In this chapter, basics of human skin biology
are described and different types of common skin lesions are briefly reviewed. The new
imaging system which is called digital dermoscopy and different clinical diagnostic
method are introduced. Also the two important dermoscopy structures, and pigment
network are defined and explained. Lastly, the current computer-aided diagnostic systems
have been reviewed and evaluated.
CHAPTER 1 Introduction
28
Chapter 4: This chapter presents pre-processing and segmentation to melanoma detection.
It has been observed that dermoscopy images often contain objects and unrelated elements,
which results in loss of accuracy as well as an increase in computational time. Thus, it
requires some pre-processing steps to facilitate the segmentation process by using different
types of filters to remove unwanted objects or artefacts. Segmentation is an important step
in the medical image analysis and classification for radiological evaluation or computer–
added diagnoses. Information is also provided about using the segmentation to partition the
input images into a number of meaningful and non-overlapped regions based on some
features that help in distinguishing between image regions (object and background).
Chapter 5: Over the past two decades Machine Learning has become one of the backbones
of information technology, although usually hidden, part of our life. With the ever
increasing amounts of data becoming available there is good reason to believe that smart
data analysis will become even more universal as a necessary part for technological
progress. The purpose of this chapter is to provide the person who reads with an overview
over the massive range of applications which have at their heart a machine learning
problem and to bring some degree of instruction to the farm of problems. After that, thesis
will discuss some basic tools from statistics and probability theory, since many machine
learning problems must be expressed to become open to solving. Finally, thesis will outline
a set of equally basic effective algorithms to solve an important problem, namely that of
classification. A discussion of more problems and a detailed analysis will follow in later
chapters of the thesis.
This chapter includes a melanoma detection model using Gabor wavelet to interest point
detection. This method displays the characteristics of certain cells in the visual cortex (the
outer layer of an internal organ) of some creatures which can be approved by these filters.
Discrete Wavelet Transform (DWT) can be used to decomposing the image into non-
overlapping sub bands [26], the mean and standard deviation of the decomposed image
portions are calculated [27]. Support Vector Machine (SVMs) as classifier directly
determine the decision boundaries in the training step and the method can also provide
good generalization in high-dimensional input spaces [28]. A hybrid technique of swarm-
based support vector machine (SVM) is introduced for melanoma detection using the
CHAPTER 1 Introduction
29
pathology images as inputs. In this technique, a particle swarm optimization (PSO) is
proposed to optimize the SVM to detect melanoma.
Chapter 6: These chapter used techniques were aiming to be able to provide
recommendation for nonspecialized users. However the variations of diagnosis are
adequacy large and there is a lack of detail of the test methods. One of the commercial
products on the market has an accuracy rate of 92%. These products is a difficult system,
taking high quality Epiluminescence Light Microscopy (ELM) images and using advanced
image analysis techniques to extract a number of features for classification which it makes
it unsuitable for a normal person to use. The traditional imaging is just a recording of what
the human eye can see by digital camera while the Dermoscopy known as
Epiluminescence Light Microscopy (ELM) images needs an experienced professional to
get the required image. This experiment presents a melanoma detection model employing a
hybrid of particle swarm optimization (PSO), wavelet Gabor (WLG), and SVM, to
improve the performance of the SVM by finding the optimal SVM parameters. This
chapter is organized in 6 Experiment Based on Digital and pathological images.
Chapter 7: This chapter presents an early detection include different stages of pre-
processing, segmentation, feature extraction and selection, and classification. Since the
output of each stage is the input of the next stage, all stages have an important role to
escape misdiagnosis. The thesis outcome results are very interesting; it is one of the best
results in this type of research as per table 6. 12 described sensitivity and specificity in
chapter 6.6.1. results and discussion. This optimal parameter could perform well. Also this
chapter include discussion and conclusion.
1.7. Publications Presented During the Doctoral Research
The doctoral research publication presented in journals and conferences are as follows:
Journal paper:
(i) “Wavelet and Curvelet Analysis for Automatic Identification of Melanoma Based
on Neural Network Classification”, (2013) in International Journal of Computer
Information Systems and Industrial Management (IJCISIM), Volume 5 ( 2013) pp.
606-614 [29].
CHAPTER 1 Introduction
30
Conference papers:
(i) “A Hybrid System for Skin Lesion Detection: Based on Gabor Wavelet and
Support Vector Machine” The model is submitted to (CISP_BMEI 2014): 7th
International Congress on Image and Signal Processing, 7th International
Conference on Biomedical Engineering and Informatics, Dalian, China 14-16
October 2014 [15].
(ii) “Automatic Recognition of Melanoma Using Support Vector Machines: A Study
Based on Wavelet, Curvelet and Colour Features”. The model is presented in
(IAICT 2014): International Conference on Industrial Automation, Information and
Communications Technology, Bali, Indonesia, Aug 28-30, 2014 [30].
(iii)“Novel feature extraction methodology based on histopathalogical images and
subsequent classification by Support Vector Machine”. The model is presented by
Md Khaled et al. (2014), in (ICCVIA2014) International Conference on Computer
Vision & Image Analysis, Ras Al Khaimah, UAE 25-27 March 2014 [31].
(iv) “Classification of Malignant Melanoma and Benign Nevi from Skin Lesions Based
on Support Vector Machine” The model is presented by Md Khaled et al. (2013), in
Fifth International Conference on Computational Intelligence, Modelling and
Simulation, Seoul, South Korea, 24-26 September [32].
(v) “The Automatic Identification of Melanoma by Wavelet and Curvelet Analysis:
Study Based on Neural Network Classification”. This model is presented by Md
Khaled et al. (2011), in 11th International Conference on Hybrid Intelligent
Systems (HIS), December 5-8, Malacca, Malaysia [33].
(vi) “Segmentation of skin cancer images based on gradient vector flow (GVF) Snake”.
The model is presented by Md Khaled et al. (2011), International Conference on
Mechatronics and Automation, August 7 - 10, Beijing, China [34].
CHAPTER 1 Introduction
31
This chapter introduction to skin cancer (melanoma) includes: research aims; problem
description; thesis objectives; the methodology used in the thesis; contribution of the
doctoral research; structure and summary of the dissertation; followed by seven doctoral
research publications one presented in journal and six presented in conferences.
CHAPTER 2 Human Skin Biology and Cancer.
32
CHAPTER 2
Human Skin Biology and Cancer.
2.1. Human Skin Biology
This chapter will discuss the structure and functions of the human tissues and cells. It takes
into account the understanding of the study of histology and how microscopes function. It
also includes the various staining procedures used to prepare tissues to make them visible
through the microscope[35].
Histology is the study of the microscopic anatomy of cells and tissues of humans,
plants and animals. It is commonly performed by examining cells and tissues through
sectioning, staining and examination under a light microscope or electron microscope.
Histology is an essential tool of biology and medicine [36]. The study of changes in cell
anatomy is known as histopathology. A sound knowledge of the normal structure of skin is
essential for an understanding of pathology. (more detailed information in appendix A)
2.2. The Cell (Structure and Function)
The cell is the functional unit of all living organisms. The simplest living organisms are
unicellular, consisting of a single cell such as bacteria and algae. More complex organisms
known as Eukaryotes have membranes surrounding their intracellular organelles.
Eukaryotes consist of a defined nucleus and an extracellular matrix. Multicellular
organisms, such as humans, display a large diversity in functional and morphological
specialization. Specialization of cell structure and function is known as cellular
differentiation [37].
Human cells mainly consist of a nucleus and cytoplasm. The cytoplasm contains a number
of organelles of which each has a defined function. The nucleus may be considered as
being the largest organelle. Figure 2.1 shows cell structure.
CHAPTER 2 Human Skin Biology and Cancer.
33
Figure 2.1 shows, Cells structure; source: www.getting-in.com
Nearly all animal and plant cells contain the following parts:• The nucleus: controls the activities that occur in the cell.
• Cytoplasm: the substance in which most chemical reactions take place.
• Cell Membrane: the barrier which holds the cell parts together and controls what
enters and leaves the cell.
• Mitochondria: where the majority of energy is produced through respiration.
• Ribosomes: where protein synthesis takes place.
Cellular replication is a vital process for growth, repair and development of unicellular
organisms which develop into more complex multicellular organisms. In eukaryotes the
cell cycle [38], as shown in Figure 2.2, is composed of four separate stages; G1, S, G2 and
M. The two (G) stages are the gap stages where the cells conduct a series of checks prior to
proceeding to the next stage. The synthesis (S) stage is where DNA replication occurs;
these three stages combined form a period known as interphase, where the cell grows and
produces proteins necessary for the cell division. The cell subsequently advances to the
mitosis (M) stage, where a cell undergoes coordinated nuclear division resulting in the
formation of two genetically identical daughter cells [39] [40].
CHAPTER 2 Human Skin Biology and Cancer.
34
Figure 2.2 The eukaryotic cell cycle, G Phases: growing, S phase: synthesis/ DNA replication, G2 Phase: Growing and preparing for Mitosis, M mitosis.
The cellular progression of the life cycle from one stage to the next is reliant upon passing
through a variety of checkpoints. Each checkpoint consists of specialized proteins to
determine if the necessary conditions exist within a cell for its progress through the cycle.
If the cellular conditions are inadequate when passing through the checkpoints, progression
will cease and the cell will remain in a dormant state known as the (G0) stage as seen in
Figure 2.3. Checkpoint errors may have significantly destructive consequences on cellular
division including cell death or mutations causing unrestrained growth i.e. cancer [41].
The human body is made up of cells, for example, when you have sunburn and your skin
peels, you are shedding skin "cells." In the center of each cell is an area called the nucleus,
and human chromosomes are located inside the nucleus of the cell. A chromosome is a
structure that contains your genes that determine your traits, such as eye color and blood
type[42].
The usual number of chromosomes inside every human’s cell is 46 total chromosomes, or
23 pairs. One half of the chromosomes is inherited (one member of each pair) from the
biological mother and the other half (the matching member of each pair) from the
biological father.
CHAPTER 2 Human Skin Biology and Cancer.
35
Figure 2.3 The cell cycle; G0 terminally differentiated cell G1 gap phase 1 G2 gap phase 2 M mitotic phase S synthesis phase
Source: Wheatear’s Functional Histology: A Text and Colour Atlas, 5th Ed
2.3. Haematoxylin and Eosin (H & E)
Haematoxylin and Eosin is the most popular staining techniques in histology and medical
diagnosis laboratories. Haematoxylin is the basic dye component which stains nucleic
acids blue/purple, including nuclei, ribosomes and endoplasmic reticulum. Due to their
high content of DNA and RNA they have a high affinity for the dye. In contrast, Eosin is
Figure 2.4 H&E-stained section of skin. High magnification (40xs) view. Haematoxylin stains the nuclei of cells blue to bluish-purple, and eosin stains other cellular elements in
the tissues from pink to red [43].
the acidic dye component which stains basic structures red/pink, such as basic cytoplasmic
proteins within the cytoplasm as shown in Figure 2.4.
2.4. Skin Cell Structure
The skin is the largest organ in the human body and is a highly organized structure
consisting of three main layers, called the epidermis, the dermis and the hypodermis as
CHAPTER 2 Human Skin Biology and Cancer.
36
displayed in figure 2.5. The skin has three main functions represented in the protection,
sensation and thermoregulation functions that are explained next.
One of the most important functions of the skin is its action as a barrier, providing
protection against a great variety of external forces. It protects against mechanical impact
and pressure e.g. friction, frequently occurring on the soles of feet and the palms of the
hands, where skin structure is slightly modified and thickened to resist such friction. It also
provides protection and acts as a barrier to water, chemicals and invasion by various
microorganisms. An example of this is microorganisms such as bacteria and fungi can live
on the skin but unless the skin is damaged, cannot penetrate into underlying tissue.
The skin is the largest sensory organ in the body. It consists of a wide range of different
receptors and nerve cells to detect and relay changes in the external environment and thus
to fulfill different roles such as for touch, temperature regulation, pain and pressure.
Interestingly due to regional variation in structure, skin which is most in contact with its
environment contains the most receptors. Damage to nerve cells is known as neuropathy
causing loss of sensation in affected areas.
The skin can act as an organ regulating the effects of several aspects of physiology. Body
temperature can be regulated via adipose tissue providing insulation, and sweat glands
secreting sweat as illustrated in Figure 2.5. The subsequent evaporation of sweat has a
cooling effect. Which are controlled by the hair erector muscles which are attached to each
individual hair follicle. Furthermore peripheral circulation responds to changes in flow of
heat and responds to certain requirement e.g. to cool the body. Heat can be lost by
increasing blood flow through the vascular networks on the surface of the skin onto the
skin surface.
CHAPTER 2 Human Skin Biology and Cancer.
37
Figure 2.5 Displayed Cross Section & the main layers of Human Skin
Source: nurrashidah2204.blogspot.com
2.5. The Layers of the Skin
The skin cell structure has three main layers as shown in figure 2.5[44]. These layers are
Epidermis, Dermis, and Hypodermis and are discussed below.
The external surface of the skin is composed of a renewable stratified squamous epithelium
with a surface layer of keratin. Keratin is a protein in direct contact with the external
environment.
The Dermis is composed of fibro-collagenous and elastic tissue; this layer consists of
blood vessels, lymphatic vessels, sensory receptors, nerves, touch and heat
mechanoreceptors, sweat glands, hair follicles and apocrine glands.
The hypodermis is the deepest layer of skin and is predominantly composed of adipose
tissue. This layer contains the larger vessels which supply and drain the dermis.
CHAPTER 2 Human Skin Biology and Cancer.
38
2.6. Melanocyte
A melanocyte cell produces the protective skin-darkening pigment
melanin. Birds and mammals possess these pigment cells, which are found mainly in the
epidermis although they also exist elsewhere as for example in the matrix of the hair.
Melanocytes could be branched or dendritic and their dendrites are used to transfer
pigment granules to adjacent epidermal cells.
2.7. Physical Properties of Human Skin
Figure 2.6 shows how the human skin color is quite variable around the world. It ranges
from a very dark brown among some Africans, Australian Aborigines, and Melanesians to
a near yellowish pink among some Northern Europeans. There are no people who actually
have true black, white, red, or yellow skin. These are commonly used color terms that do
not reflect biological reality[45] .
Figure 2.6. Skin Color Distribution around the World
Source: www.gbhealthwatch.com
As displayed in figure 2.6, the skin color is due primarily to the presence of a pigment
called melanin, which is controlled by at least 6 genes. The number and size of melanin
particles differs in each individual. These two variables are most important in determining
skin color in comparison to the percentages of the different kinds of melanin. In lighter
skin, color is also affected by red cells in blood flowing close to the skin. The presence of
fat under the skin or carotene can also be contributing factors. Also, hair color is another
factor controlled by the presence of melanin[45].
CHAPTER 2 Human Skin Biology and Cancer.
39
As shown in figure 2.7(b), melanin is normally located in the epidermis, or outer skin
layer. These melanocyte cells have photosensitive receptors, similar to those in the eye,
that detect ultraviolet radiation[46] from the sun and other sources. In response, they
produce melanin within a few hours of exposure. Melanin acts as a protective biological
shield against ultraviolet radiation which helps prevent sunburn damage which may cause
DNA changes and, subsequently, several kinds of malignant skin cancers. Ultraviolet
radiation reaching the earth usually increases in summer and decreases in winter. The
skin's ability to tan in summertime is adjusted to this seasonal change. Tanning is
a) Cross section of human skin, Source: anthro.palomar.edu
b) Structure of Epidermis Source: silver-botanicals.com
Figure 2.7. a) Cross section of human skin, b) Structure of Epidermis.
CHAPTER 2 Human Skin Biology and Cancer.
40
primarily an increase in the number and size of melanin granules due to the stimulation of
ultraviolet radiation.
It would be harmful if melanin acted as a complete shield. A certain amount of shortwave
ultraviolet radiation (UVR) [47] must penetrate the outer skin layer in order for the body to
produce vitamin D. Approximately 90% of this vitamin in people normally is synthesized
in their skin and the kidneys from a cholesterol-like precursor chemical with the help of
ultraviolet radiation. The remaining 10% comes from foods such as fatty fish and egg
yolks.
2.8. Human Skin
The skin is the largest organ of the body. It provides protection against heat, sunlight,
injury, and infection. Skin also helps in controlling body temperature and stores water, fat,
and vitamin D.
There are two main kinds of human skin:
• One is Glabrous skin (non-hairy skin), found on the palms and soles; it is grooved
on its surface by continuously alternating ridges and sulci, which has individually
unique configurations known as dermatoglyphics. It is characterized by a thick
epidermis which is divided into several well-marked layers, including a compact
stratum cornea, by the presence of encapsulated sense organs within the dermis,
and by a lack of hair follicles and sebaceous glands.
• “The second type is the hair-bearing skin which has both hair follicles and
sebaceous glands but lacks encapsulated sense organs. There is also wide variation
in skin types between different body sites.
Skin is mainly composed of three primary layers as displayed on Figure 2.7(a)
• The epidermis, which provides waterproofing and serves as a barrier to infection
[48].
• The dermis, which serves as a location for the appendages of skin
CHAPTER 2 Human Skin Biology and Cancer.
41
• The hypodermis (subcutaneous adipose layer)
Cancer is a disease of the body's cells that in the end destroys living organs, which can lead
to death. Cancer is the term used to describe a lot of different diseases including malignant
tumors, leukemia (a disorder of white blood cells), sarcoma of the bones, Hodgkin's
disease and non-Hodgkin's lymphoma (affecting the lymph nodes) in which uncontrolled
cell growth threatens the rest of the body. There are more than 100 different types of
cancer, but the most common five types are: bowel cancer, breast, prostate, melanoma
and lung cancer (excluding non-melanoma skin cancer) and these form 60% of all cases.
In Australia, the most common cancers in men are prostate, colorectal, lung cancers and
melanoma, but in women the most common is breast cancer, colorectal cancer, melanoma
and lung cancer.
Like all body tissues, the skin is made of tiny building blocks called cells. These cells can
sometimes become cancerous, from exposure to ultraviolet (UV) radiation[49]. The top
layer of skin contains three different types of cells that can be affected: basal cell cancer,
squamous cell cancer and melanoma, as shown in figure 2.8.
Australia has the highest incidence of skin cancer in the world,[49-51] [52] [53] and non-
melanoma skin cancer (NMSC) is the most commonly diagnosed cancer in Australia[54].
Skin cancer in Australia’s most expensive cancer, with more than $264 million (9% of the
total costs for cancer) spent on diagnosis and treatment in 2000–01[55].
In 2009 the review article[56] concludes that: In the Australian patient population, the
whole body skin examination is essential to avoid missing concurrent lesions. Patients with
concurrent lesions have a high risk of developing future NMSC
CHAPTER 2 Human Skin Biology and Cancer.
42
Figure 2.8. The Layers of Skin - the Epidermis and Dermis (At the top, the close up shows
a squamous cell, basal cell, and melanocyte). Source: www.uchospitals.edu
.
Figure 2.9(a), shows the most common and least dangerous skin cancer Basal Cell
Carcinoma (BCC). It appears as a lump or scaly area and can be red, pale or pearly in
color. It grows slowly – usually on the head, neck or upper torso – and can become
ulcerated as it grows.
Figure2.9- a) Basal Cell Carcinoma (BCC), b) Squamous Cell Carcinoma (SCC)
Figure 2.9(b), shows the Squamous Cell Carcinoma (SCC) cancer; it grows over a period
of weeks or months and may spread to other parts of the body if not treated promptly. It
CHAPTER 2 Human Skin Biology and Cancer.
43
occurs most often (but not only) on areas exposed to the sun. This can include the head,
neck, hands and forearms. This cancer looks like thickened, red, scaly spots.
Melanoma is the most dangerous type of skin cancer; it results from sun damage and other
causes. A skin biopsy can identify melanoma [57] Melanoma grows quickly and develops
over weeks to months. If caught early, it is usually curable, but if it spreads to other parts
of the body, it can be very difficult to cure. Melanoma appears as a new spot or as an
existing spot, freckle or mole that changes color, size or shape. It usually has an irregular,
smudgy outline and is often more than one color.
Melanoma as shown in Figure 2.10 can occur anywhere on the skin, even on the soles of
the feet. Melanocytes in the eye, respiratory system, nervous system, even on the cortex of
the brain and mucous membranes (e.g. lining of the mouth and nasal passages) can also
become cancerous. These types of melanoma are rare; 90% of the cases appear on skin.
Figure 2.10 – Melanoma
Melanoma is malignant tumor of melanocytes [58]. Melanocytes produce the dark
pigment, melanin, which is responsible for the color of skin. These cells predominantly
occur in skin, but are also found in other parts of the body, including the bowel and
the eye. Melanoma can originate in any part of the body that contains melanocytes. It
causes the majority (75%) of deaths related to skin cancer[59]. The majority of malignant
melanomas are caused by heavy sun exposure in white-skinned populations [60]. Incidence
rates are highest by far in Australia/New Zealand (37 per 100,000 in 2008), where it is the
third most common cancer in both males and females, accounting for one in nine (around
11% in 2008) of the total cases [61]as displayed in figure 2.11.
CHAPTER 2 Human Skin Biology and Cancer.
44
Figure 2.11: Malignant Melanoma, World Age-Standardised Incidence Rates, World Regions, 2008 Estimates.
At later stages, the mole may itch, ulcerate or bleed[56]. Anyone who has more than 100
moles is at greater risk for melanoma. It is so important to get to know your skin very well
and to recognize any changes in the moles on your body. Have a check-up with the
physician, and then contact the specialist for early diagnosis[56].
All cancers are caused by damage to the DNA inside cells. This damage can be inherited in
the form of genetic mutations, but in most cases, it builds up over a person's lifetime and is
caused by factors in their environment. DNA damage causes the cell to grow out of
control, leading to a tumor. Melanoma is usually caused by damage from UV light from
the sun, but UV light from sunbeds can also contribute to the disease[62].
A number of rare mutations, which often run in families, are known to greatly increase
one’s susceptibility to melanoma. Several different genes have been identified as
increasing the risk of developing melanoma. Some rare genes have a relatively high risk of
causing melanoma; some more common genes, such as a gene called “MC1R” that causes
red hair, have a relatively lower elevated risk. Genetic testing can be used to determine
whether a person has one of the currently known mutations.
The earliest stage of melanoma starts when the melanocytes begin to grow out of control.
Melanocytes are found between the outer layer of the skin (the epidermis) and the next
layer (the dermis). This early stage of the disease is called the radial growth phase, and the
tumor is less than 1 mm thick. Because the cancer cells have not yet reached the blood
vessels lower down in the skin, it is very unlikely that this early-stage cancer will spread to
other parts of the body. If the melanoma is detected at this stage, then it can usually be
CHAPTER 2 Human Skin Biology and Cancer.
45
completely removed with surgery. When the tumor cells start to move in a different
direction — vertically into the epidermis and into the papillary dermis — the behavior of
the cells changes dramatically[63].
Visual diagnosis of melanomas is still the most common method employed by health
professionals[64]. Moles that are irregular in color or shape are often treated as candidates
for melanoma. The diagnosis of melanoma requires experience, as early stages may look
identical to harmless moles or not have any color at all. People with a personal or family
history of skin cancer or of dysplastic nevus syndrome (multiple atypical moles) should see
a dermatologist at least once a year to be sure they are not developing melanoma. There is
no blood test for detecting melanomas.
Melanoma is divided into the following types [65] [66]:
• Lentigomaligna melanoma: Begins as a large freckle in an area of skin that gets a
lot of sun exposure, such as the face and upper body as displayed in figure 2.12. It
may grow slowly and superficially over many years, later forming lumps as it
grows deeper into the skin.
Figure 2.12. Lentigomaligna Melanoma.
Source: melanomaknowmore.com
• Superficial spreading melanoma: Forms 70% of the cases, it grows in the top layer
(epidermis). It is dangerous when it invades into the lower layer (dermis) and it is
dark brown or black in color as displayed in figure 2.13.
CHAPTER 2 Human Skin Biology and Cancer.
46
Figure 2.13. Superficial Spreading Melanoma.
Source: melanomaknowmore.com
• Acral lentiginous melanoma: Most commonly found on the palms of the hands and
soles of the feet or under the nails. More common in people with darker skin.
• Mucosal melanoma.
• Nodular melanoma: Forms 15% of the cases, it is very dark brownish black or
black in color but can be pink or red and forms a lump on the skin surface; also it
invades deeper into the skin as displayed in figure 2.14.
Figure 2.14. Nodular Melanoma.
Source: melanomaknowmore.com
• Polypoid melanoma.
• Desmoplastic melanoma.
• Amelanotic melanoma.
• Soft-tissue melanoma.
• Unclassified: up to 5% of cases.
CHAPTER 2 Human Skin Biology and Cancer.
47
Also of importance are the "Clark level" and "Bristow’s depth”, which refer to the
microscopic depth of tumor invasion[67].
Figure 2.15 represents melanoma stages [68] and the survival rates in 5 year:
Stage 0: Melanoma in situ (Clark Level I), 99.9% survival
Stage I / II: Invasive melanoma, 89–95% survival.
Stage II: High risk melanoma, 45–79% survival
Stage III: Regional metastasis, 24–70% survival
Stage IV: Distant metastasis, 7–19% survival
Figure 2.15.Stages of Cancer Development and Metastasis.
Source: http://www.cancervic.org.au/about-cancer/advanced-cancer
Clark's level is a related staging system, used in conjunction with Bristow’s depth, which
describes the level of anatomical invasion of the melanoma in the skin[69].
CHAPTER 2 Human Skin Biology and Cancer.
48
Five anatomical levels are recognized, and higher levels have worsening prognostic
implications. These levels are:
• Level 1 : Melanoma confined to the epidermis (melanoma in situ)
• Level 2 : Invasion into the papillary dermis
• Level 3 : Invasion to the junction of the papillary and reticular dermis
• Level 4 : Invasion into the reticular dermis
• Level 5 : Invasion into the subcutaneous fat[69].
In medicine, Breslow's depth was used as a prognostic factor in melanoma of the skin. It is
a description of how deeply tumor cells have invaded. Currently, the standard Breslow's
depth has been replaced by the AJCC depth. As shown in table 2.1, Breslow's depth was
originally divided into 5 stages[70].
Table 2.1: Breslow's depth
Stage Depth
Stage I less or equal to 0.75mm
Stage II 0.75 mm - 1.5mm
Stage III 1.51 mm - 2.25mm
Stage IV 2.25 mm - 3.0mm
Stage V greater than 3.0 mm
The above studies showed that depth was a continuous variable correlating with prognosis.
However, for staging purposes table 2.2 shows the most recent AJCC guidelines use
cutoffs of 1 mm, 2 mm, and 4 mm to divide patients into stages.
Table 2.2: Survival figures from British Association of Dermatologist Guidelines 2002
Tumor Depth Approximate 5 year survival
<1 mm 95-100%
1 - 2 mm 80-96%
2.1 - 4 mm 60-75%
>4 mm 50%
CHAPTER 2 Human Skin Biology and Cancer.
49
Advances in the early detection and treatment of cancer has resulted in a dramatic increase
in the number of cancer survivors,[5] and the numbers are expected to grow considerably
[71-75]. More than 60% of cancer patients will survive more than five years after
diagnosis. Also the survival rate for many common cancers has increased by more than
30 % in the past two decades [75]. The chart on table 3.3, shows incidence (2009), and
morality (2010) of melanoma of the skin of Australia [76]. The chart on table 3.4, shows
observed incidence (2009) and morality (2010) of non-melanoma of the skin and estimated
for 2012 in Australia.
It was estimated that more than 120,700 Australians were diagnosed with cancer in the
year 2012, excluding basal and squamous cell carcinoma of the skin[77]. Males made up
more than half (56%) of these cases. In 2012 the most commonly reported cancers were
prostate cancer, bowel cancer, breast cancer, melanoma of the skin and lung cancer.
Cancer accounted for about 3 in 10 deaths in Australia in 2010, were the death rate was
more than 42,800. For all cancers combined, the age-standardised mortality rate decreased
by 17% from 210 per 100,000 in 1991 to 174 per 100,000 in 2010[77]. As indicated in
Figure 2.16 five-year relative survivals from selected cancers by remoteness area,
Australia. 2006-2010[77].
Figure 2.16: Five-year relative survival from selected cancers by remoteness area, Australia. 2006-2010.
CHAPTER 2 Human Skin Biology and Cancer.
50
It was found that cancer outcomes were different when it came to Aboriginal and Torres
Strait Islander status, remoteness area and socioeconomic status. Indigenous Australians
experienced higher incidence and mortality rates than non-Indigenous Australians for all
cancers combined. Incidence rates and survival were lower for people living in remote
areas compared with those in major cities, while mortality rates rose with increasing
remoteness. Incidence and mortality rates rose and survival from all cancers fell as a
person’s socioeconomic status decreased[77] . Table 2.3: displays the incidence (2009) and
morality (2010) of melanoma of the skin and estimated for 2012. Also Table 2.4: displays
the incidence (2009) and morality (2010) of non-melanoma of the skin and estimated for
2012.
This chapter ‘Human Skin Biology and Cancer’ includes basics of human skin biology; the
cell structure and function; hematoxylin and eosin. This chapter also describes staining
techniques in histology and medical diagnosis laboratories. In the subsequent sections of
this chapter also includes the layers of the skin; melanocyte cell; physical properties of
human skin; human skin structure and cancer types like melanoma staging as well as non-
melanoma; the chapter also includes study about health behavior of cancer survivors in
Australia.
CHAPTER 2 Human Skin Biology and Cancer.
51
Table 2.3: Observed incidence (2009), and morality (2010) of melanoma of the skin and estimated for 2012.
Source: AIHW Australian Cancer Database 2009, AIHW National Mortality Database
CHAPTER 2 Human Skin Biology and Cancer.
52
Table 2.4: Observed incidence (2009), and morality (2010) of non-melanoma of the skin and estimated for 2012
Source: AIHW Australian Cancer Database 2009, AIHW National Mortality Database.
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
53
CHAPTER 3
Image Processing and Automatic Skin Cancer Diagnosing
3.1. Introduction
The occurrence of skin cancer (malignant melanoma [MM]) has increased dramatically
over the past few decades[41]. Australia is one of many countries in which skin cancer is
more widely spread in comparison to other types of cancer[78]. The Health Benefits and
Risks of Sun Exposure researchers which reported that solar radiation is the main cause of
skin cancers. However, it also is a main source of vitamin D for humans[49]. Researchers
found that, primary prevention and early detection continue to be of paramount importance
in addressing the public health threat of skin cancer[79] [12].
The diagnosis of a doctor is based on visual inspection and may be reliable, although this
procedure takes a lot of time and effort. These routines can be automated rather than the
previous procedure saving doctors lots of time and would ensure that the diagnoses are
more accurate[80]. There are other good means of storing diagnostic information alongside
computer means. These means will help assist further investigations or the creation of new
methods of diagnosis based on a vision system.
3.2. Vision System & Structure of the Human Eye
Electromagnetic radiation in the optical band is generated from the visual environment we
view which enters the visual system of the eye and is detected in the sensitive cells of the
retina. The activities start in the retina, where the signals from neighboring receivers are
compared and a coded message dispatched on the optic nerves to the cortex, behind our
ears.
An excellent account of human visual perception may be found in[81]. The spatial
characteristics of our visual system have been proposed as a nonlinear model in[82] [83].
Although the eyes can detect tranquility and static images, they are essentially motion
detectors. The eyes are capable of identification of static objects and can establish spatial
relationships among the various objects and regions in a static scene. Their basic
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
54
functioning depends on comparison of stimuli from neighboring cells, which results in
interpretation of motion. When observing a static scene, the eyes perform small repetitive
motions called saccades that move edges past receptors. The perceptual recognition and
interpretation aspects of our vision, however, take place in our brain. The objects and
different regions in a scene are recognized in our brain from the edges or boundaries that
encapsulate the objects or the regions inside the scene.
The maximum information about the object is embedded along these edges or boundaries.
The process of recognition is a result of learning that takes place in our neural
organization. The orientation of lines and the directions of movements are also used in the
process of object recognition.
Light from an object moves through space and reaches our eyes enabling us to see. Signals
are then sent to our brain, and our brain deciphers the information in order to detect the
appearance, location and movement of the objects we are seeing. This process would not
be possible if the presence of light was not available. Without light, sight would not be
possible. The human eye is the organ which allows us to sense sight allowing us to learn
more about the surrounding world, more than any of our other five senses. The eyeball is
set in a protective cone-shaped cavity in the skull called the orbit or socket and measures
approximately one inch in diameter. As Figure 3.1 illustrates, soft, fatty tissues surround
the orbit protecting the eye and enabling it to turn easily.
In general the eyes of all human and animals look like simple cameras in that the lens of
the eye forms an inverted image of objects in front of it and projects it onto the sensitive
retina, which corresponds to the film in a camera. As figure 3.2 shown , the front of the eye
mainly contains an elaborate array of structures which are mainly concerned with the
refraction (i.e. bending) of light rays and bringing them into focus on the retina. The
structure most directly involved is the lens. The eye changes light rays into electrical
signals then sends them to the brain, which interprets these electrical signals as visual
images via the optic nerve as in figure 3.3.
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
55
Figure 3.1 - Show the function of the eye.
Source: http://cdn2.hubspot.net/hub/60407/file-350482438-jpg/images/Master_Eye_logo_short_tag_line_revised_10-13-2013-resized-600.jpg
Figure 3.2 - Distance of vision.
Source:hyperphysics.phy-astr.gsu.edu/hbase/vision/accom.html
Figure 3.3 - Structure of the Eye.
Source:drpion.be/en/bouw-van-het-oog.htm
3.3. Image Processing & Human skin Imaging
Utilizing the concept of the biological vision system, an image processing system been
established. An image processing system contains many components. To start to process
digital images, we need to understand how the image processing operates. The brightness
variation of an image should be described using its histogram, (making the image brighter
or dimmer). We shall also consider thresholding techniques that turn an image from grey
level to binary (two values representation). These are called single point operations. As a
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
56
result we see how the statistical operations reduce noise in the image, which benefits the
feature extraction techniques which need to be considered later. As such, these basic
operations are usually for pre-processing to improve display quality.
Digital image processing is based on the conversion of a continuous image field to
equivalent digital form. A digital image is a detached function defined over a rectangular
grid (lattice) representing the characteristics of the objects being imaged[84]. In 2-D
images, each grid element, or “pixel”, is defined by a location and a value representing a
characteristic of the object at that location. The intensity and brightness are also factors
which influence the process. Images, in which each pixel is assigned a single value, as in
the above example are called scalar-valued images. Images in which each pixel is
represented by a vector are called vector-valued or multispectral images. Typical examples
of multispectral images in medicine and biology are red green blue (RGB) color images.
As mentioned in[84], in the Matlab Image processing toolbox, there are four types of
images: Indexed, Intensity, Binary, and RGB.
• Indexed image: consists of a data matrix and a color map matrix.
• Intensity image: consists of a single data matrix, each element of which denotes the
intensity of the corresponding pixel (medical imaging).
• Binary image: is stored as a logical array. (i.e., each pixel is turned on or off), and
is represented by a byte.
• RGB image: sometimes referred to as a true- color image, is stored in Matlab as an
m-by-n-by-3 data array that defines red, green, and blue color components for each
individual pixel.
Image compression techniques help to reduce the redundancies in raw image data in order
to reduce the storage and communication bandwidth.
The Automatic Visual Inspection System improve the productivity and the quality of the
product in manufacturing and allied industries[85]. There are many applications of image
processing; some are listed below:
• Remotely Sensed Scene Interpretation: Information regarding the natural resources,
such as agricultural, hydrological, mineral, forest and geological resources, etc., can
be extracted based on remotely sensed image analysis[86] [87].
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
57
• Biomedical Imaging Techniques: Various types of imaging devices are used
extensively for the purpose of medical diagnosis[88] [89].
• Defense surveillance: Application of image processing techniques in defense
surveillance is an important area of study. There is a continuous need for
monitoring the land and oceans using aerial surveillance techniques.
• Content-Based Image Retrieval: The features of a digital image (such as shape,
texture, color and topology of the objects, etc.) can be used as index keys for search
and retrieval of pictorial information from a large image database. Retrieval of
images based on such image contents is popularly called the content-based image
retrieval[90] [91].
• Moving-Object Tracking: Tracking of moving objects, for measuring motion
parameters and obtaining a visual record of the moving object, is an important area
of application in image processing[6] [92]. Generally there are three different
approaches to object tracking:
1. Recognition-based tracking.
2. Motion-based tracking.
3. Image and Video Compression: This active application area in image
processing[91]. Development of compression technologies for image and
video continues to play an important role for success of multimedia
communication and applications.
There are many algorithms in machine vision which can provide reliable automatic object
detection, recognition and classification in an independent environment for understanding
two- and three-dimensional objects in a digital image [93],[94],[95],[96]. There are a
number of questions concerning vision, such as: (i) what are the goals and limitations?
(ii) What type of algorithm or set of algorithms is required to affect vision? (iii) What are
the Implications for the process, given the types of hardware that might be available?
(iv) What are the levels of representation required to achieve vision? The levels of
representation are dependent on what type of segmentation can and/or should be applied to
an image. We are required to consider (a) how to best segment an image and (b) the form
this segmentation should take. Segmentation is the first step in automatic image analysis
and pattern recognition. In these cases, we need appropriate techniques of refining the
images to get better visual quality, which are free from aberrations and noises. Image
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
58
enhancement, filtering, and restoration have been some of the important applications of
image processing since the early days of the field[97] [98] [84].
Image analysis also includes topics such as feature extraction, shape representation, and
object recognition and classification. In this research, we will cover some of the
fundamental topics that have found widespread applications in medical image processing.
Medical digital images are pictures of distributions of physical attributes captured by an
image acquisition system. Medical images may show anatomy including the pathological
variation of anatomy if the measured value is related to it or physiology when the
distribution of substances is traced. Examples of medical imaging include X-ray imaging,
CT, MRI, nuclear imaging, ultrasound imaging, photography, and microscopic images
[99].
With respect to digital imaging, four major and several minor imaging techniques meet
these requirements. The major techniques are as follows:
• X-ray imaging measures the absorption of short wave electromagnetic waves; this
is known to vary between different tissues.
• Magnetic resonance imaging measures the density and molecular binding of
selected atoms (most notably hydrogen which is abundant in the human body),
which varies with tissue type, molecular composition, and functional status.
• Ultrasound imaging captures reflections at the boundaries between and within
tissues with different acoustic impedance.
• Nuclear imaging measures the distribution of radioactive tracer material
administered to the subject through the blood flow. It measures function in the
human body.
Other imaging techniques include EEG and MEG imaging, microscopy, and photography.
All the techniques have in common that an approximate mapping is known between the
diagnostic question, which was the reason for making the image and the measurement
value that is depicted. This can be very helpful when selecting an analysis technique. If
bones need to be detected in an x-ray CT slice, a good first guess would be to select a
thresholding technique with a high threshold because it is known that x-ray attenuation in
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
59
bone is higher than in soft tissues and fluids[99]. The first major component is a camera
that captures the images.
The sensors which are used in most of the cameras are either Charge Coupled Device
(CCD) or CMOS sensors. The CCD camera comprises a very large number of very small
photo diodes, called photosets. The electric charges which are accumulated at each cell in
the image are transported and are recorded after appropriate Analogue to Digital
Conversion (ADC).
In a situation where bright sunlight is present, the aperture which is located behind the
camera lens needs to be small as little light is required, whereas on cloudy days more light
is needed to create an image so the aperture should be enlarged. This is identical to the
functioning of our eyes. The shutter speed gives a measure of the amount of time during
which the light passes through the aperture. The opening and closing of the shutter depends
on how much light is available. The distance between the focal planes included the focal
point and the focal length was measured from the optical ‘centre’ of the lens and the
surface of the sensor array is known as the focal length of a digital camera. Focal length is
the critical information in selecting the amount of required magnification which is desired
from the camera as shown in figure 3.4.
A digital camera can capture images in various resolutions, (low, medium or high
resolution). The cameras that we normally use can produce about 16 million colors, i.e., at
each pixel we can have one of 16 million colors. The spatial resolution of an image refers
to the image size in pixels, which corresponds to the size of the CCD array in the camera.
Figure 3.4 Focal length defined [When parallel rays of light strike a lens focused at
infinity, they converge to a point called the focal point. The focal length of the lens is then defined as the distance from the middle of the lens to its focal point.]
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
60
A digital camera can capture colors in many ways. A practical solution is based on the
concept of color interpolation or demo-sailing, which is a more economical way to record
the three primary colors of an image. In this method, we permanently place only one type
of filter over each individual photo-site. Usually the sensor placements are carried out in
accordance to a pattern. The most popular pattern is called the Bayer's pattern[100], where
each pixel is indicated by only one color-red, blue, or green pixel. It is possible to make
very accurate guesses about the missing color component in each pixel location by a
method called color interpolation or demosaicing [101] [102].
In high-quality cameras, however, three different sensors with three filters are used and
light is directed to the different sensors by using a beam splitter. Each sensor responds only
to a small wavelength band of color. Thus the camera captures each of the three colors at
each pixel location.
Dermatology is often described as a visual specialty wherein a majority of diagnoses can
be made by visual inspection of the skin. Diagnosis of skin disease in dermatology is
largely non-invasive. The physician diagnosis is based on the anatomic distribution, color,
configuration, and visible surface changes of a lesion. In some cases, a skin biopsy is
performed which again offers the opportunity for a microscopic visual examination of the
lesion in question, but there exist limitations in the assessment of depth and size of skin
lesions as well as internal features of superficial lesions. Such limitations result in the need
for an objective non-invasive means of assessing the skin.
Digital dermatoscopic images firstly have to be parameterized for automatic classification.
Extensive study of skin nature has to be done in literature to parameterize it.
Imaging modalities with major relevance to skin imaging include various modification of
electromagnetic wave imaging (Optical, Infrared, Tera Hertz, and Nuclear Magnetic
Resonance), acoustical wave imaging and mechanical wave imaging. Usually tomographic
images i.e. 2D cross sectional images are acquired with medical imaging systems.
Depending on the spatial orientation of the cross section, these images depict information
about the tissue over depth or any other direction. Three dimensional tissue volumes are
usually imaged by acquiring a stack of consecutive 2D images[103].
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
61
The different illumination method called epiluminescence can be used in order to get the
image from deeper skin layers. The light is directed straight in to these layers and reflected
back giving more information about consistency of light absorbers in these layers. This
method of illumination improves diagnosis accuracy up to 10 %.
Another interesting solution for getting more information from skin is by using multi
spectral photography, which uses narrow frequency bands of light illumination. Those
images give information about consistency and concentration of absorbers and reflectors
in the skin. When those photos are made with a range of light waves, the reflectance
frequency characteristics of skin can be calculated and compared to normal skin. We can
calculate the reflectance frequency characteristics of skin and compare to normal skin.
Some of the methods based upon the above classification of skin imaging are described in
coming sections[104].
In the last two decades an increasing rate of malignant melanoma has been observed
[105, 106]. Because of a limited range of effective treatments after the spread of cancer (
melanoma), the best treatment currently is still early diagnosis and punctual surgical
excision of the primary cancer[105] Dermoscopy (also known as epiluminescence
microscopy, dermoscopy, and amplified surface microscopy) is a non-invasive and in live
method, that has been reported to be a useful tool for the early recognition of malignant
melanoma [107] [108] [109].
The performance of dermoscopy has been investigated by many authors. Its use increases
diagnostic accuracy between 5% and 30% over clinical visual inspection depending on
the type of skin lesion and experience of the physician [110-113]. This was confirmed by
recent evidence-based publications and based on statistics of the literature [107]. This
section is a review of the principles of dermoscopy as well as recent technological
developments.
The study of histology is carried out using a variety of microscopes, in order to visualise
the structure of body tissues. Pathologists use photomicrographs, captured by the light
microscope (LM) (color images) and the electron microscope (EM) (black and white
images), which differ in optical resolution and available magnification. Resolution
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
62
practically refers to the capacity of an optical system to reveal details within a specimen.
EM images are therefore said to display cell and tissue ultrastructure. There are two types
of electron microscope producing two types of images, either two or three dimensional.
The scanning EM which produces a three-dimensional (3D) image is restricted to surface
structure and the transmission EM utilizes an electron beam which must pass through the
specimen to form an image. (more detailed information in appendix B)
The strengths of LM and EM differ yet complement one another very effectively. Large
areas of a specimen can be observed under a light microscope. Identification of cell and
tissue features can be accentuated through utilizing a wide range of staining methods. The
electron microscope has superior resolution and magnification permitting observation of
many features and intracellular structures which simply cannot be seen via the light
microscope.
Historically the transmission electron microscope (TEM) was the first type of Electron
Microscope to be developed and is patterned similarly to the light transmission microscope
except that a focused beam of electrons is used instead of light.
The advancements in science led to a desire to observe intracellular structures in greater
detail and interior structure of organic cells e.g. the nucleus. Light microscopes are limited
by the physics of light consequently the electron microscope was developed. This provided
required much greater magnification of 10,000 X plus, which was possible using the
current optical microscope. Electron microscopes are scientific instruments that examine
objects on a very fine scale by using a beam of energetic electrons[114]. The similarities in
the overall design and principal features of the optical microscope (OM), the transmission
electron microscope (TEM) and the scanning electron microscope (SEM) are shown in
Figure 3.5.
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
63
Figure 3.5. Displayed, similarities of overall design in the principal features of an optical microscope, a transmission electron microscope and a scanning electron microscope.
Source: www.nslc.wustl.edu
Confocal laser scanning microscopy (CLSM) is a new optical microscopic technique,
which offers significant advantages over conventional microscopy. CLSM is microscopy
of optical sections. Light, which is emitted from regions other than the focal plane, is cut
off by introducing a diaphragm in the beam path. The result is an optical "slice", which
shows more details because the blurring from out of focus haze disappears. It has been
repeatedly used in experimental, but also in diagnostic dermatopathology.
The "in vivo” confocal microscopy, applied directly to the intact skin provides details of
living cells in the superficial layers comparable to that of fixed and stained tissue. While
the extent of its future applications is hard to predict, its potential for applications in
dermatology appears very considerable, particularly for studies of fixed or living tissues,
where it is desirable to obtain clear images many micrometers below the surface of the
tissue under examination [115].
The confocal microscope is elementally comprised of a coherent scanning laser light
source, condenser lens, objective lens, and detector. A laser light point source illuminates a
small region within the skin and is scanned across the skin. Light reflected from this focal
point propagates back to the detector through a pinhole aperture after a single collision
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
64
within the tissue, allowing the observation of reflections from single scattering events.
Light from the in focus plane is emanating from above and below the focus plane
minimally passes through the pinhole to the detector, resulting in images of high resolution
and contrast. The term “confocal” stems from this design in which the pinhole in front of
the light source and the pinhole in front of the detector are in optically conjugate focal
planes. Figure 3.6 illustrates the principle of confocal microscopy.
Figure 3.6. Principle of CSLM. Note the same optical path is used for the detector and the source. Optics are used to direct the light towards the detector
Confocal laser scanning microscopy is a novel technique that enables us to non-invasively
investigate skin at the cellular level and observe dynamic changes that may occur over a
period of time. It was shown that the incident light wavelength in CSLM systems affects
signal intensity and structural details at the cellular level. Systems equipped with shorter
wavelengths may provide improved resolution in images, while those with longer
wavelengths can image deeper into the skin. When comparing systems of the same design
and only differing in incident wavelength, the 785 nm system produces images with greater
signal intensity and contrast in structural detail compared to the 830 nm system. It is
peculiar that imaging human skin with 785 and 830 nm wavelengths produces intensity
profiles that appear biphasic, while imaging at 405 nm result in intensity profiles that
follow a simple exponential decay[116].
Capturing cellular changes over time with CSLM is of great significance as it enables the
study of the physiology of living skin tissue. Skin is the body’s largest functioning organ,
consisting of millions of cells that are in constant activity whether it is proliferation,
pigmentation, repair, etc. Capturing the same cells over time and observing their activity
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
65
provides powerful diagnostic information. By studying healthy skin tissue and
understanding how it reacts and functions under average conditions, gives a significant
insight into the cells of healthy tissue. Skin tissues can be better characterized and the
differentiation between healthy and unhealthy tissue made. Capturing physiological and
morphological changes in skin due to aging, treatment and wounding is only the beginning
of the fundamental understanding of dynamic cellular changes using CSLM. Having a
strong foundation in the knowledge of normal function and activity of healthy tissue is
essential to the study and understanding of diseased tissue.
There are many publications on the subject of the differential diagnosis of pigmented
lesions of the skin. The 5 algorithms most commonly used are pattern analysis[117] [113]
[118]; the ABCD rule of dermoscopy [41] [119] [119]; the 7-point checklist [105] [117]
[120] ; the Menses method[121], [122], [123]; and the revised pattern analysis [124]. The
common two-step procedure for the classification of pigmented lesions of the skin is
illustrated in Figure 3.7. The first step is the differentiation between a melanocytic and a
non melanocytic lesion. For this decision, the algorithm in Figure 3.8 is used. The
outcomes image results are displayed on Figures from 3.9 to 3.15.
Figure 3.7. Two-step procedure for the classification of pigmented skin lesions. Adapted from Argenziano.[122]
To detect melanomas (and increase survival rates), it is recommended to learn what they
look like ("ABCDE" as displayed below), to be aware of moles and check for changes
(shape, size, color, itching or bleeding) and to show any suspicious moles to a doctor with
an interest and skills in skin malignancy [125].
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
66
Figure 3.8. Algorithm for the determination of melanocytic versus non melanocytic lesions according to the proposition of the Board of the Consensus Netmeeting. Adapted
from Argenziano.[122]
Figure 3.9. A, Macroscopic picture of a superficial spreading malignant melanoma (Breslow thickness 0.52 mm; Clark level II). B, Dermoscopy of A shows (atypical)
pigment network and branched streaks and can therefore be considered a melanocytic lesion.
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
67
Figure 3.10. A, Macroscopic picture of a blue nevus. B, Dermoscopy of A shows steel-blue areas (no pigment network, no aggregated globules, and no branched streaks).
Figure 3.11. A, Macroscopic picture of a seborrheic keratosis. B, Dermoscopy of A shows comedolike openings (a), multiple milia-like cysts (b), and fissures (c).
Figure 3.12. A, Macroscopic picture of a seborrheic keratosis. B, Dermoscopy of A shows comedolike openings and multiple milia-like cysts.
Figure 3.13. A, Macroscopic picture of a basal cell carcinoma. B, Dermoscopy of A shows maple lifelike areas, ovoid nests, and arborized telangiectasia.
Figure 3.14. A, Macroscopic picture of a basal cell carcinoma. B, Dermoscopy of A shows multiple spoke wheel areas.
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
68
Figure 3.15. A, Macroscopic picture of an angioma. B, Dermoscopy of A shows red lagoons.
Anyone who has more than 100 moles is at greater risk for melanoma. It is so important to
know the skin very well and to recognize any changes in the moles on the body. The
ABCDE rules of melanoma can help in such case, if a person see one or more of this signs;
they should check-up with the physician, and then contact the specialist for early
diagnosis.[56]. The Asymmetry Border Color Diameter Evolution (ABCDE) popular
method for remembering the signs and symptoms of melanoma is as displayed in figure
3.16. It is based on check the skin lesion as below:
• Asymmetrical skin lesion: shows the two halves will not match.
• Border: shows the border edges may be scalloped or notched.
• Color: A number of different shades of brown, tan or black could appear. A
melanoma may also become red, blue or some other color.
• Diameter: shows that melanomas are larger in diameter than the size of the eraser
on your pencil (1/4 inch or 6 mm).
Figure 3.16. Asymmetry Border Color Diameter Evolution (ABCD-E) rule for Diagnosis of Melanoma
Source: www.webmd.com
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
69
Evolving: shows the change in size, shape, color, evasion, and bleeding, itching or crusting
these points to danger
The pictures on Figure 3.17, shows the typical normal moles and melanomas these are
useful in showing people the differences as a start point.[126]
The ABCD rule is introduced here and can be easily learned and rapidly calculated and has
been shown to be a reliable method providing a more objective and reproducible diagnosis
of melanoma. The semi-quantitative ABCD rule represents the second step of a two-step
procedure that was originally proposed by [127] and who later modified the rule to be
applied, at least following Stoltz’ instructions [128].
To improve diagnostic accuracy the ABCD rule of lesion screening is widely used based
on asymmetry (A), border (B), color (C), and differential structure (D). Measuring criteria
have to be assessed semi-quantitatively. Then, each of the criteria has to be multiplied by a
given weight factor yielding a total dermoscopy value (TDV) which results from the
calculation as per rule (1).
TDV = (A score x 1.3) + (B score x 0.1) + (C score x 0.5) + (D score x 0.5) …. (1)
This score contributes to the differentiation between benign and malignant lesions:
1.00 – 4.75 – Benign skin lesion
4.75 – 5.45 – Suspicious lesion
More than 5.45 – Highly suspicious for melanoma.
Nodular melanomas do not fulfill these criteria, having their own mnemonic, "EFG":
• Elevated: the lesion is raised above the surrounding skin.
• Firm: the nodule is solid to the touch.
• Growing: the nodule is increasing in size.
If the mole is malignant, the mole and an area around it need excision. Elliptical excisional
biopsies may remove the tumor (figure 3.18(a)); this is followed by histological analysis
and Bristow scoring refers to figure 3.18(b). Punch biopsies are contraindicated in
suspected melanomas, for fear of seeding tumor cells and hastening the spread of the
malignant cells.
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
70
Benign Malignant
Figure 3.17 - Normal moles and Melanomas [Benign and Malignant].
Source: www.skinbychar.com
Total body photography, which involves photographic documentation of as much body
surface as possible, is often used during follow-up of high-risk patients. The technique has
been reported to enable early detection and provides a cost-effective approach (being
possible with the use of any digital camera), but its efficacy has been questioned due to its
inability to detect macroscopic changes[64]. The diagnosis method should be used in
conjunction with (and not as a replacement for) dermoscopy imaging, with a combination
of both methods appearing to give extremely high rates of detection.
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
71
Figure 3.18: Showed: a) Melanoma in skin biopsy with H&E Stain (“This case may represent superficial spreading melanoma.”), b) Lymph node with almost complete replacement by metastatic melanoma. The brown pigment is focal deposition of melanin.
In 1998 Argenziano [4] and his colleagues described a 7- point checklist based on the
analysis of 342 pigmented skin lesions. In[45],[129],[130] they distinguish 3 major criteria
and 4 minor criteria. Each major criterion has a score of 2 points while each minor
criterion has a score of 1 point. A minimum total Score of 3 is required for the diagnosis of
malignant melanoma.
According to Menses[131] for diagnosing melanoma, both of the following negative
features must not be found: a single color (tan, dark brown, gray, black, blue, and red, but
white is not considered) and ‘‘point and axial symmetry of pigmentation’’ (refers to pattern
symmetry around any axis through the center of the lesion). This does not require the
lesion to have symmetry of shape. Additionally, at least one positive feature must be
found.
In figure 3.19, the three-point checklist was developed for inexperienced dermoscopists.
This is meant to encourage them to use dermoscopy by teaching them a simplified
algorithm for the evaluation of pigmented skin lesions. General practice physicians can
easily be taught dermoscopy and the three-point checklist, which can be utilized as a
screening tool to determine whether a given pigmented lesion should be further evaluated
by a dermatologist. The three-point checklist [132] features three criteria that are found to
be particularly important in the differentiation between malignant and benign pigmented
skin lesions by participants in the Consensus Net Meeting on Dermoscopy [122]. In the
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
72
three-point checklist the presence of any two of the three criteria: asymmetry; atypical
pigment network; and blue-white structures indicates that the lesion under investigation
may be a melanoma.
Figure 3.19. Diagnosis of Melanoma using Three-point Checklist
Source: www.dermoscopy.org
There is need for appropriate techniques of refining the images to get better visual quality,
free from aberrations and noise. Some methods that need to be considered for describing
skin lesion properties for diagnosing processing are:
• Asymmetry of the lesion border.
Transition of the pigmentation from the skin lesion to the surrounding skin. (Edge
Gradient)
• Colour distribution of the skin lesion including the blue-white veil.
We found that there are a number of various imaging methods of skin lesions.
The simplest visualization method is photography. This method gives only the top layer
skin image.
Dermoscopy Imaging Techniques
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
73
In order to get deeper layer image oil immersion is used. It reduces reflections of surface
and brightens the image off the epidermis – the second skin layer. Better results are
reached when photos are made with polarized light source from the top layer of skin as
shown in Figure 3.20. Then it is easier to estimate the lesion structures like dots, globules,
nets that are the major indicators to melanoma diagnosis.
Figure 3.20 - Example of pigmented skin lesion. Left: Traditional imaging technique. Right: Dermoscopy imaging technique.
Another interesting solution of getting more information from skin is by using multi-
spectral photography. It uses a narrow frequency band of light illumination. Those images
give information about consistency and concentration of absorbers and reflectors in the
skin. The idea is that a different pigment of skin absorbs different light wave frequencies,
determining the color of our skin. When those photos are made with a range of light waves,
we can calculate the reflectance frequency characteristics of skin. These photos can be
used to compare with normal skin characteristics and diagnostic decisions can be made
about skin pigment consistency.
Ultrasound visualization is usually used to measure depth of melanoma and if there is no
melanoma practically there are no differences. When a doctor diagnoses melanoma, then
he uses high frequency ultrasound (over 30 MHz) to measure penetration depth in order to
make the correct cut during surgery.
It is a non-invasive diagnostic technique used in dermatology for the in vivo observation of
pigmented skin lesions. Dermoscopy is an imaging technique that provides a more direct
link between biology and distinct visual characteristics. It is known as epiluminescence
light microscopy (ELM).
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
74
This diagnostic tool allows for a better visualization of surface and subsurface structures
and thus permits the recognition of morphologic structures which are not visible to the
naked eye. This technique has opened a new dimension for analysing the clinical
morphologic features of pigmented skin lesions.
Dermoscopy images are much more detailed than conventional macroscopic (clinical)
images. Macroscopic images are merely images of skin lesions seen under a magnifying
glass. Figure 3.21 shows a comparison of a macroscopic and a dermatoscopic image of the
same skin lesion.
Figure 3.21. (a) Macroscopic Image of a Lesion (b) Dermatoscopic Image of the same Lesion. Source: www.jle.com
Studies have shown that dermatoscopic diagnosis has an accuracy of 75% to 97% while for
macroscopic diagnosis this range is 65% to 80% [36] [38]. However, it has also been
noticed that dermoscopy may actually lower the diagnostic accuracy in the hands of
inexperienced dermatologists [38]. Therefore, due to the lack of reproducibility and
subjectivity of human interpretation, the development of computerized techniques is of
utmost importance [133].
Work reported in [134], where it found Matlab very handy tool which allows for automatic
detection of skin irregularities and which was used to calculate lesion properties like
asymmetry of shape, or border irregularities, which can help in detecting melanoma. There
are a lot of investigations in this work; finally, the author provides his own source code to
let researchers trial they own data. Figure 3.22. Showed my data results using the trial
code.
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
75
[A] Image Center of mass.
Original image
Filtered Image
Black and White
Traced
[B] Image process.
Figure 3.22- Show the results of skin lesion boundary tracing algorithm, my data experiment shown [a] Image center of mass, [b] Image process.
Automated border detection is one of the most important steps in this procedure as the
accuracy of the subsequent steps essentially depends on the accuracy of this step.
Unsupervised approach to border detection in dermoscopy skin lesion images is based on a
modified version of the Japan Society of Engineering geology (JSEG) algorithm. This
method presented achievements of both fast and accurate border detection in dermoscopy
images[135] [136]. Some of the digital processing techniques are contour detection and
image segmentation.
Some studies have combined image processing with knowledge-based approaches in order
to achieve better results in interpreting images by presenting a general methodology
developed for computer interpretation and integration of medical images. Plain anatomical
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
76
models, as well as other domain knowledge, are used to facilitate the feature extraction and
fusion of images from different modalities.
Segmentation is to break an image up into meaningful segments. Segmented images
improve image analysis especially for image querying and retrieval [137]. There are many
different approaches to segment images; this section will describe some techniques
commonly used with segmentation.
Image segmentation via the use of data clustering is a very effective technique[137]. When
it is used with the appropriate distance and probability distribution functions,
segments/clusters of an image can be properly distinguished. Often, the most difficult task
in data clustering in general is determining the optimal number of clusters[137].
Specifying too few clusters will generate incomplete image segments, and specifying too
many will generate redundant segments. Figure 3.23 shows an example of the cluster
threshold problem.
Generally, grey scale images are easier to segment compared to color images. This is due
to the fact that grey scale images consist of only one color component which is the grey
level, ranging from 0 to 255. Color images are generally stored in the RGB color space.
The RGB format is not a strong model for calculating distances between color pixels since
it is not a uniform color space, and as such, it will not perform well for color
segmentation[137]. Calculating distance measures for color pixels requires the use of a
uniform color space. Usually, the color spaces are preferred for color segmentation
distance calculations.
Texture segmentation is classified into two main categories, supervised and unsupervised.
In cases where the different types of textures are small, supervised segmentation is used
where the a-priori knowledge of the textures is known [138]. In cases where the types of
textures are large or the assumptions of the texture cannot be made, unsupervised texture
segmentation must be used.
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
77
Figure 3.23- From left to right: the original image (source image from: http://ijplugins.sourceforge.net/plugins/clustering/), clustered image using 2 clusters
(poor), clustered image using 3 clusters (close but one key cluster is missing), clustered image using 4 clusters (optimal), and clustered image using 10 clusters (artifacts are
noticeable).
Texture segmentation involves applying the texture filters on the image and then using
segmentation methods such as the clustering technique to segment an image into textured
as such segments.
Many texture filters have been implemented in the past. Some of the commonly used filters
are the Gabor filters[139] and the wavelet-based decomposition/ transformations [140].
K-means is one of the common approaches for data clustering. It attempts to subdivide a
region of pixels into smaller regions based on the region’s pixel mean value, and by doing
so, pixels that are closer to the mean value will be grouped together[141]. The process
continues until it reaches a stage where nothing has changed. The naive approach to
finding the distance between pixel values in the feature space is to use the Euclidean
Distance. Although it is a faster way to determine distances between pixel values, it
assumes that clusters are shaped in hyper-spheres only, which is quite restrictive.
The Euclidean distance in equation 3.1:
3.1
Where is a matrix of the difference between the two data vectors and [142].
The Mahalanobis Distance method models the distance between color pixels more
accurately compared to the Euclidean distance model [141]. It calculates the distances
using the variance values of a cluster’s data set. That is, the distance between color pixels
is calculated using the variance values between each color component.
The Mahalanobis distance in equation 3.2:
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
78
3.2
Where is a matrix of the difference between the two data vectors and with
respect to cluster , and the covariance matrix for cluster [142]
Fuzzy set theory application in image processing is used in areas like enhancement,
segmentation, filtering, edge detection, content-based image retrieval, pattern recognition
and clustering [143].
Fuzzy logic concepts and application in digital image processing have become apparent.
Fuzzy set theory is thus useful in handling computer vision and image-processing
applications. Fuzzy logic has three main stages, namely, image fuzzification, modification
of membership values, and, if necessary, image defuzzification as shown in figure 3.24.
The main part of fuzzy image processing in the membership plane is shown in steps of
fuzzy image process in figure 3.25.
Figure 3.24- Structure of fuzzy image processing.
Figure 3.25- Steps of fuzzy image processing.
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
79
Another modification of the K-means clustering is the Fuzzy C-means (FCM) clustering
algorithm. Data points in the non-fuzzy clustering methods can only belong to a single
cluster [144]. In the FCM approach, data points can belong to more than one cluster based
on some predefined membership rules that determine the degree of membership within
each cluster for each data point. In the basic form of K-means clustering, each data points’
distance from its cluster center contributes equally to its centroid location. This makes the
cluster centroid values weak to outliers. The FCM method reduces the force of outliers by
reducing the influence of each outlier for those clusters that are least likely to have
generated them [145].
Clustering can be improved by using the Expectation Maximization (EM) algorithm[141].
It computes the probabilities of cluster memberships by maximizing the log likelihood of
the data that were generated by the Gaussian mixture model. Other possible probability
distributions are the Normal, log-normal, and the Poisson distributions, all of which
provide a statistical measure of the image information. Another important note to mention
is the EM method can be applied to continuous or categorized data. Categorization of data
will allow for weighting of data; thus areas of an image that are considered less important
can be successfully separated from the key areas.
The gradient vector flow (GVF) technique utilizes the snake algorithm to detect the
borders of the lesion areas of an image [146]. It uses luminance blurring (for correction of
color saturation) to remove noise from a dermatology image during its initialization
process. The snake border is then drawn based on the outline of the blurred lesion. It then
successively applies the blurring technique to deform the snake border, which would
eventually conform to the shape of the lesion. However, this technique is not accurate as
weaker outlines/edges in the original image would fade away after the initial blurring step,
thus the snake border will not conform to the original shape of the lesion.
Use of a double image buffer to rectify this problem has improved the overall accuracy of
the snake border, but the error rates are still not acceptable.
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
80
Markov Random Field (MRF) models have been applied to the field of digital image
restoration of degraded images [147]. The image restoration process is closely related to
image segmentation and boundary detection. MRF models are a powerful technique and
consist of two layers, a region labeling component and a feature modeling component, with
a constant weighting factor to combine the two. Many researchers have mentioned that the
improved MRF model is able to produce accurate results in image segmentation.[148]
Since the 1960s, the clinical characteristics of melanoma, its histopathology and its
biological basis have been the subject of intense study at pigmented lesion clinics in North
America, Europe, and Australia [149]. The purpose of this future work is to bring to the
attention of anatomic pathologists the essential characteristics of the pathology report for
primary cutaneous melanoma in the modern era.
The researchers in [150] [151] recognized the importance of anatomic compartment of
attack to the melanoma. In 1970 the researcher in [70] described the use of an optical
micrometer to specifically measure primary tumor thickness. These two parameters
became the recognized prognostic variables available to histo-pathologists in predicting the
biological behavior of melanoma [152]. Throughout the 1980s, advances in biology and
pathology of melanoma made it apparent that tumors undergo specific growth phases. In
particular, the radial growth phase was shown to be unable to generate metastatic
generating metastatic events unless a worsening was seen at the primary site. Further work
carried out by Massachusetts General Hospital and the University of Pennsylvania
established that the existence of microscopic satellites and brisk mitotic activity could
modify the behavior of a tumor.[149]
Successful treatment of melanoma, by surgical removal depends upon the early detection
of the lesion. The challenge is to diagnose and remove all malignant lesions at an early
stage and at the same time minimize the unnecessary removal of benign lesions. However,
the differentiation of early melanoma from other non-malignant pigmented skin lesions is
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
81
not easy even for the experienced dermatologists. In several cases primary care physicians
underestimated melanoma in its early stage [153].
Visual inspection with the naked eye has a relatively low sensitivity and reliability in
detecting early melanoma. In this context, several non-invasive diagnostic modalities, such
as dermoscopy, total body photography, and reflectance confocal microscopy (RCM), have
emerged in recent years that are aimed at increasing diagnostic accuracy. The main
developments in this field are the integration of dermoscopy and digital photography into
clinical practice.
Dermoscopy (also known as dermoscopy or epiluminescence microscopy) helps in the
visualization of anatomical structures within the epidermis and papillary dermis that cannot
be visualised by the naked eye. This is achieved with the use of a hand held dermatoscope
to magnify the skin surface and reduce the refraction of light by the corneal layer. It has
been reported that dermoscopy training for primary care physicians can improve their
ability to correctly refer individuals with suspicious lesions and decrease the rate of
excision or referral in benign skin lesions.
Pathologists are using the microscopic to perform histo-pathological examination and
provide diagnostic information based on their observations. Figure 3.26 shows manual
segmentation performed by pathologists using the ‘Aperio Image Scope’ software [154],
but there are many problems associated with manual segmentation, being time-consuming
and depends on the skill and training of the pathologist doing the analysis. Therefore
automatic segmentation needed to bring some kind of consistent accuracy within the
process. On the technical report in [155], the automated cancer diagnosis consists of three
main computational steps: pre-processing, feature extraction, and diagnosis. The aim of the
diagnosis step is (i) to distinguish benignity and malignancy or (ii) to classify different
malignancy levels by making use of extracted features. This step uses statistical analysis of
the features and machine learning algorithms to reach a decision.
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
82
Figure 3.26. Manual Segmentation by trained pathologists using the ‘Aperio Image Scope’ software [156].
Malignant melanomas are the most serious form of skin cancer accounting for the majority
of skin cancer related deaths. Histo-pathological images of skin tissues are analysed for
detecting various types of melanomas. The automatic analysis of these images can greatly
facilitate the diagnosis task for dermato-pathologists. The first and primary step in
automatic histo-pathological image analysis is to accurately segment the images into
dermal and epidermal layers along with segmenting other tissue structures such as nests
and melanocytic cells which indicate the existence of cancer.
An efficient automatic segmentation procedure has been proposed for histo-pathological
images of skin tissue. The proposed technique is based on the Orientation Sensitive Fuzzy
C-Means (OS-FCM) algorithm along with other refinement techniques [154].
The radial growth phase growing in a horizontal array as single cells and small nests is
predominantly in an intra-epidermal location.[149] The radial growth phase is incapable
biologically of generating metastatic events, save for specific and uncommon
circumstances. In this example, there is pagetoid spread of fully transformed malignant
epithelioid melanocytes within the epidermis, see (Figure 3.27) the basement membrane
zone in concert with a level II component in the papillary dermis (i.e. Clark level II). Note
that the dermal component comprises small nests of neoplastic melanocytes (red arrow
figure.3.27) with a similar cytology to those seen above the basement membrane zone
which separates them from the overlying epidermal component (black arrow figure 3.28).
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
83
The dermal nests of radial growth phase melanoma are smaller than those seen in the
overlying epidermal component.
Clark in [157] described the early vertical growth phase and distinguished it from the
radial growth phase by criteria that included, most importantly, the existence of a dominant
nest within the papillary dermis. This expansile nest is larger than any nest within the
epidermis or the surrounding dermis (Figure 3.28). Clark also suggested that the nodule
should have 25–50 cells to aid in its identification, and that the cells comprising this nodule
are cytological separate from those of the intraepidermal component by virtue of their
different characteristics, such as shape, size, nuclear or cytoplasmic features, or the
presence or absence of pigment.
The vertical growth phase means the point at which the melanoma becomes biologically
capable of producing metastatic events. Note that there is a main able to expand dermal
nest of neoplastic melanocytes (arrow) which is larger than any of the overlying the quiz at
the dermal–epidermal junction or within the epidermis.
Figure 3.27 Showed the Radial growth phase melanoma.
Since the 1970s melanomas have been histologically sub-classified as belonging to four
major groups: superficial spreading, lentigo malign, acral lentiginous, and nodular
melanoma.[151], [158], [159] Some observers suggest that the relatively greater
survivorship seen in lentigo maligna melanoma vs. the other histologic subtypes is
reflective of smaller thickness at time of diagnosis.[157], [160], [161]
The lifetime risk for developing invasive disease in a lesion of lentigo malign has been
estimated to be only in the 5% range. [162] The early work of Fiddler[163] suggests that
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
84
those are the blood vessels capable of accommodating the large tumor bolus conducive to
the establishment of the metastatic seed and so suggest that this is a critical factor in the
mortality of any given acral melanoma.
In the accumulated experience of several investigators over a long period of time, tumor
thickness as measured by ocular micrometer has emerged as the most powerful predictor of
outcome in primary cutaneous melanoma [70], [164], [157], [8], [165], [166], [167], [168,
169], [170], [171], [7], [172]. The Breslow measurement is taken from the epidermal
surface or, in the event that the surface is ulcerated, from the base of the ulcer, and is made
with a calibrated ocular micrometer. [70] Long clinical follow-up indicates that melanoma
thickness is a continuous variable associated with dropout of survivors beyond 10 years.
The Breslow measurement is the most important means of prognosticating mucosal
melanomas that lack the anatomic compartmentalization seen in the skin. [173] In most
prognostic studies, the measured depth emerges as the most powerful independent factor
for prediction of lymph node metastasis and survival.[174]
A review of the microscope slides of the primary tumours for clinical Stage I melanoma
discovered that primary lesions displayed invasion with "microscope satellites” [175].
characterized by reticular dermal and/or sub-circular nodules of tumor greater than
0.05mm in width separated from the main vertical growth phase component (Figure 3.28),
are associated with lymph node metastasis, and with decreased disease-free and overall
survival [157],[7] ,[176], [177], [178], [179], Clark et al.[10]
Figure 3.28 showed the Vertical growth phase melanoma
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
85
found a reduction of actuarial 8-year survival from [180] to 40% when satellites were
identified. (Figure 3.29) It has been suggested that this is a difficult criterion to assess, as
tumor tongues radiating from the main vertical growth phase nodule may mimic satellites
due to artifacts of sectioning.[181]
Figure 3.29 Microscopic satellites, (a). Shows neoplastic group is discontinuous (arrow) from the overlying vertical growth phase component. Shows the melanocytes must have a
malignant cytomorphology (b).
The presence or absence of blood vessel and lymphatic invasion should be reported (Figure
3.30). Regression of over 75% of the tumor volume of a melanoma is considered a bad
prognostic sign. Regression is characterized by haphazard, pattern-less fibrosis,
accompanied by vertically oriented, ectatic blood vessels (straight arrow), melanophages
(curved arrow) and patchy lymphocytic infiltration of the stroma. In complete, as opposed
to partial regression, there is a total absence of neoplastic melanocytes in the dermis and in
the overlying epidermis.
• Cellular atypia: cancerous melanocytes have atypical cellular features compared to
normal melanocytes, such as:
Increased nuclear: cytoplasm ratio
Found migrating throughout the layers
• Nests: atypical melanocytes are usually found in nests suggesting uncontrolled
growth
• Lymphocytic infiltrate: the presence of immune cells suggests the attempt of an
immune response to a malignancy
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
86
Figure 3.30 Regression: Regression of over 75% of the tumor volume of a melanoma is considered a bad prognostic sign.
The above discussion showed the essential elements of the pathology report for primary
melanoma should include all of the statistically proven analytical variables.
Dermoscopy is a non-spreading method that allows in vivo evaluation of colors and
microstructures of the epidermis, the dermo-epidermal junction, and the papillary dermis
not visible to the naked eye. The pigmented skin lesion is covered with liquid (usually oil
or alcohol) and examined under a specific optical system during a dermoscopy assessment.
Recently by using polarized light in dermoscopy with LED light with polarization,
immersion liquid is no longer necessary, and some of these instruments do not need direct
skin contact. Researchers have reported excellent agreement for most dermoscopy colors,
with the exception of blue-white veil and pink (red) color when comparing non-polarized
and polarized light [182]. They concluded that the polarized light improves the
visualization of red areas and vessels, especially the latter with non-contact dermoscopy
[182].
Figure 3.31(a) shows an image of a lesion under the clinical (naked eye) view and 3.31(b)
shows the same lesion under a dermoscopy with oil immersion. Important features are
marked in the image. This methodology is reserved for experienced clinicians because of
the complexity involved. See figure 3.32, for more details.
Figure 3.31: a) An image of a lesion under clinical view (naked eye). b) Shows the same lesion under a dermoscopy with oil immersion.
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
87
The four clinical diagnostic methods, which have received the largest interest among
dermoscopy users, will be briefly reviewed. Initially, pattern analysis was the first
dermoscopy method presented for diagnostics of pigmented skin lesions [183]. Firstly the
International Dermoscopy Society (IDS) modified and refined the pattern analysis method.
Then the other three methods with the largest impacts are ABCD rule, Menses method and
the 7-point checklist which have been described in previous sections. All of these methods
showing similar results on sensitivity for melanoma, but specificity differed slightly in
favor of pattern analysis [184]. As a result, the accuracy of these simplified systems is
somewhat lower than the full system, pattern analysis. For more details on the dermoscopy
of pigmented skin lesions please refer to [183].
Figure 3.32: a) Colors allow the physician, to some extent, to draw conclusions about the localization of pigmented cells within the skin. Black and brown indicate pigmentation in
the epidermis, while grey and blue correspond to pigmented cells within the superficial and deep dermis, respectively [185].
Figure 3.33 shows different kinds of dermoscopy and they are divided into analogue and
digital types. Analogue types are more widely used however digital types are easier to take
and store dermoscopy images. The magnification in all dermoscopy was identical. The
authors show that in the newer dermoscopy, the image quality with regard to color and
visible differential structures are distinctly improved compared to the dermoscopy (Heine
Delta 10R) with only one light source[13].
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
88
Figure 3.33: Figures (a, b, c, d, and e) show analogue dermoscopy. All of them, except (a), are attachable to digital cameras to function as digital dermoscopy. Figures (f, g, and h) show DinoLite, Handy scope, and Dermoscopy which are modern digital dermoscopy.
Advances in computer science and applied mathematics are now allowing the basic
technology of dermoscopy to be extended to a more complex application: forecasting
whether observed spots are cancerous. Several groups are working toward computer-
assisted diagnosis of melanoma using different mathematical and analytical strategies.
Computer software can be used to library skin images and allow remote diagnosis and
reporting by a dermatologist (digital epiluminescence microscopy, teledermoscopy, mole
mapping). Mole Map NZ is such a system which is used for archiving and patient
management. If we want to develop computer-aided diagnosis systems, we should address
the important step of the pattern analysis approach which is the segmentation and analysis
of the dermoscopy structures
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
89
Solar Scan is a device priced for use by individual practitioners and developed by
Australian startup Polartechnics, with assistance from Australia’s Commonwealth
Scientific and Industrial Research Organization (CSIRO) and the Sydney Melanoma
Unit[186]. Solar Scan uses image analysis, which has turned diagnostic criteria used by
doctors into defined decision trees that lead toward or away from a decision to help general
physicians (GPs) diagnose melanoma[187]. The device works by capturing an image of a
patient’s skin spot using an object shaped like a hairdryer with a built-in surface
microscope. Image analysis software removes unnecessary things from the image like hairs
and oil bubbles and analyses the spot’s features such as its shape and color. Solar Scan
then compares the features against images of melanomas and non-melanomas in a
database, returning an advice to a GP. A record of the spot’s status can be stored in
Polartechnics’ Body Map software so that it can be rechecked another time if
necessary[187].
In figure 3.34, MelaFind is a multispectral digital dermoscopy with a specialized imaging
probe and software to assist with differentiation between early melanoma and other skin
lesions[188] [186]. A hand-held imaging device that shines light of 10 different, specific,
wavelength bands is used to collect data. Proprietary processing software is used to extract
specific features from the images [189] [190]. The software determines the edge of the
lesion and generates a 10 digital image sequence. The images are then analyzed for
wavelet maxima, asymmetry, color variation, perimeter changes, and texture changes, and
the output is a binary recommendation of whether or not to perform a biopsy [191]. The
results of a clinical trial were published in April 2008, leading to submission of the
MelaFind to the FDA and recently it has been approved for use in the United States.
We believe that with new advances in dermoscopy, the texture analysis problem should be
changed to an object recognition problem that involves identification, segmentation and
recognition of individual shapes and structures in skin lesions and this will be employed to
identify and classify specific dermoscopy structures. It is hoped that this will lead to the
development of many new approaches that can be included to increase the diagnostic
accuracy of automated systems.
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
90
Figure 3.34. The Solar Scan instrument: (a) global appearance, (b) camera, (c) user interface [192]. Source: www.medgadget.com
Chapter summary This chapter ‘Literature Review’ states that: the diagnosis of a doctor is based on visual
inspection and may be reliable, although diagnosis procedure takes a lot of time and effort.
These routines can be automated rather than the previous procedure saving doctors lots of
time and would ensure that the diagnoses are more accurate. There are other good means of
storing diagnostic information alongside computer means. These means will assist further
investigations or the creation of new methods of diagnosis based on a vision system.
This chapter includes briefly reviewed sections like basics of vision system and structure
of the human eye; image processing: human skin imaging; human skin biology; different
types of common skin lesions, the new imaging system called digital dermoscopy and
different clinical diagnostic method. Lastly the current computer aided diagnostic systems
and other commercial equipment like Solar Scan Melanoma Monitoring and MelaFind are
also reviewed.
The discussion in this chapter showed the essential elements of the pathology report for
primary melanoma including all of the statistically proven analytical variables.
• As explained before in the modern era, clinicians and oncologists must recognize
specific risk factors for metastatic disease so as to guide adjuvant chemotherapy or
immunomodulatory therapy and further surgical therapy including sentinel and
completion lymphadenectomy when necessary.
CHAPTER 3 Image Processing and Automatic Skin Cancer Diagnosing
91
• It is therefore considered essential that patients receiving melanoma diagnoses be
provided with accurate and comprehensive diagnostic assessment upon which these
novel therapeutic strategies will be based.
CHAPTER 4 Pre-processing and Segmentation
92
CHAPTER 4
Pre-processing and Segmentation
4.1. Introduction
The rate of skin cancer (malignant melanoma [MM]) has increased dramatically over the
past few decades. Researchers found that, Primary avoidance and early detection continue
to be of paramount importance in addressing the public health threat of skin cancer[193]
[79].
Doctor’s diagnosis is reliable, but this procedure takes lots of time and effort. These
routines can be automated. It could save time and could help the diagnosis to be more
accurate[80]. Besides using computerized means there are good opportunity to store
information with diagnostic information in order to use it for further investigations or
foundation of new methods of diagnosis.
The main problem is that the diagnosis is highly dependent on individual judgment, and is
just reproducible [113] [194]. Several counting systems and algorithms such as the ABCD-
E rule [119] [195, 196], the seven-point checklist[197] [198], three-point checklist[199]
and the Kenzie’s method [200] [201] have been proposed to improve the diagnostic
performance of less experienced clinicians. Although this simplification has enabled the
development of these diagnostic algorithms with good accuracy but still they showed
problems that have not yet been solved. The most important is that the purpose for which
they were designed was not achieved, because the within- and between-observer
concordance is very low, even for expert observers[119] [111] [202] [203]. Despite
extensive research in investigating the varied presentations and physical characteristics of
melanoma, the clinical diagnostic accuracy remains sub optimal. Thus, a growing interest
has developed in the last two decades in the automated analysis of digitized images
obtained by ELM techniques to assist clinicians in differentiating early melanoma from
benign skin lesions.
CHAPTER 4 Pre-processing and Segmentation
93
A digital image is a binary representation of a two-dimensional image. The digital
representation is an array of picture elements called pixels. Each of these pixels has a
numerical value which in monochromatic images represents a grey level. An image may be
defined as a two-dimensional function, where and are spatial coordinates, and the
amplitude of at any pair of coordinates is called the intensity of the image at that
point. The term grey level is used often to refer to the intensity of monochrome images.
Color images are formed by a combination of individual 2-D images. For example, in the
RBG color system, a color image consists of three (red, green, and blue) individual
component images. this is why many of the technics developed for monochrome images
can be extended to color images by processing the three component images
individually[27].
An image may be continued with respect to the and coordinates, and also in amplitude.
Converting such an image to digital form requires that the coordinates, as well as the
amplitude, be digitized. Digitizing the coordinate values is called sampling; digitizing the
amplitude values is called quantization. Those, when and the amplitude values of
are all determinate discrete quantities, we call the image a digital image.
The image types as supported in Matlab / (IPT) Image Processing Toolbox are divided into
four types of images:
• Intensity images: consist of a single data matrix, each element of which denotes the
intensity of the corresponding pixel (medical imaging).
• Binary images: stored as a logical array. (i.e. each pixel is turned on or off), and is
represented by a byte.
• Indexed images: consist of a data matrix and a couloir map matrix.
• RGB images: sometimes referred to as a true-colour image, is stored in Matlab as
an m-by-n-by-3 data array that defines red, green, and blue colour components for
each individual.
Most monochrome image processing operations are carried out using binary or intensity
images, so our focus is on these two image types.
1. Intensity images
CHAPTER 4 Pre-processing and Segmentation
94
An intensity image is a data matrix whose values had been scaled to represent intensities.
When the elements of an intensity image are of class unit8, or class unit16 they have
integer values in the range ( and ( respectively. If the image is of class
double, the values are floating – point numbers. Values of scaled, class double intensity
images are in the range by convention.
2. Binary images
Binary images have a very specific meaning in Matlab. A binary image is a logical array
of and . Thus, an array of and whose values are of data class, say, unit8 is not
considered a binary image in Matlab. A numeric array is converted to binary using
function logical. Thus, if is a numeric array consisting of and , we creating a
logical array using the statement in equation 4.1:
4.1
If contains elements other than and , use of the logical function converts all
nonzero quantities to logical and all entries with value to logical Using relational
and logical operators also creates logical arrays. To test if an array is logical we use the
islogical function, equation 4.2:
4.2
If is a logical array, this function returns a . Otherwise it returns a Logical arrays can
be converted to numeric arrays using the data class conversion functions.
Application of computational intelligence methods helps physicians as well as
dermatologists in faster data process to give better and more reliable diagnoses. Studies
related to the automated classification of pigmented skin lesion images have appeared in
the literature as early as 1987[204]. After some successful experiments on automatic
diagnostic systems for melanoma diagnosis[204] [205] [206] , utility of machine vision
and computerized analysis is getting more important every year. The importance of the
topic is obvious if we analyze the enormous quantity of researches related with the
melanoma diagnosis. Several computerized diagnostic systems have been reported in the
literature where different border detection, feature extraction, selection and classification
algorithms are used. Some researchers [205] [207] [208] reviewed and tried to critically
CHAPTER 4 Pre-processing and Segmentation
95
examine image analysis techniques for diagnosis of skin cancer and compared diagnostic
accuracy of expert’s dermoscopists with artificial intelligence and computer aided
diagnosis. More research, however, is needed to identify and reduce worries in the
automatic decision support systems to improve diagnosis accuracy.
Computer aided decision-support tools are important in medical imaging for diagnosis and
evaluation. Analytical models are used in a variety of medical domains for diagnostic and
prognostic tasks. These models are built from experience which establishes data acquired
from actual cases. The data can be pre-processed and expressed in a set of rules, such as it
is often the case in knowledge-based expert systems and accordingly serve as training data
for statistical and machine learning models.
The general approach of developing a CAD system for diagnosis of skin cancer is to find
the location of a lesion and also to determine an estimate of the probability of a disease.
The first step in this research was to establish a standard general scheme of a CAD system
for skin lesions. This scheme is shown in figure 4.1.
The inputs to the Computer Aided Systems are digital images obtained by ELM, with the
possibility to add other acquisition system such as ultra-sound or confocal microscopy etc.
As showed in figure 4.1. In the first phase a pre-processing of image is done which allows
us to reduce the ill effects and various objectives like hair, the true color Pathological
Digital Images filterd, then followed by the detection of the lesion by image segmentation
technique as shown in figure 4.2. Once the lesion is localized, different chromatic and
morphological features can be quantified and used for classification.
Differentiation of malignant melanoma images demands very fast image processing and
feature extraction & classification algorithms. A detailed research is necessary to make the
best choice and setting the benchmarks for diagnostic system development and support.
Next image processing modules consist of four section focuses on the description of the
major steps that may be involved in skin cancer diagnosis, which is connected to a medical
knowledge database can generate a computerized diagnosis, suggesting whether the lesion
is benign or malignant.
CHAPTER 4 Pre-processing and Segmentation
96
Original Pathology Digital Image
Evaluation output
Figure 4.1 Showed: Four stages CAD system for skin lesions.
Figure 4.2 true color Pathological Digital Image.
4.2. Pre-Processing Stage
The main processing step towards a complete analysis of pigmented skin lesion is to
differentiate the lesion from the healthy skin. Detection of the lesion is a difficult problem
in dermatoscopic pathology images as the transition between the lesion and the
surrounding skin is smooth and even for trained dermatologist, it is a challenge to
distinguish accurately. It has been observed that dermoscopy images often contain objects
such as rough brightness dermoscopy gel black frames, ink markings, rulers, air bubbles,
as well as essential development features that can affect border detection such as blood
CHAPTER 4 Pre-processing and Segmentation
97
vessels, hairs, and skin lines and texture. These objects and unrelated elements complicate
the border detection procedure, which results in loss of accuracy as well as an increase in
computational time. Thus, it requires some pre-processing steps to facilitate the
segmentation process by removal of unwanted objects.
Everything that might corrupt the image and consequently affect the results of image
processing must be localized and then removed, masked or replaced. It may include image
resizing, masking, cropping, hair removal, and conversion from RGB color to intensity
grey image. It is done to reduce noise and the effect of reflection objects. It is meant to
facilitate image segmentation by filtering the image and enhancing its important features.
The most straightforward way to remove these artefacts is to smooth the image using a
general purpose filter such peer group filter (PGF)[209], mean filters, median filter [210]
[211, 212], Gaussian filters[213] [214] or anisotropic diffusion filters (ADF). An
alternative solution is to use filters that treat the pixels as vectors[215]. Another notable
thing is setting of mask size proportional to the image size to manage a trade-off between
smoothing of image and blurring of edges. Regardless of taking care of all the fore
mentioned things, it is still not guaranteed to get an image free of all objects.
Many studies suggested their works, very few[216] [217] [217] discussed different aspects
of objects together but none of them have discussed all cases of objects or the factors that
complicate the detection of borders in dermoscopy images with insufficient contrast. The
contrast of image is enhanced to ensure that edges of the lesion are distinction.
Recently, it has been proposed a contrast enhancement method based on independent
histogram pursuit (IHP) [218]. An easy, yet powerful way to enhance the image contrast is
histogram stretching, a mapping of the pixel values onto (0, 255). Another very popular
technique is histogram equalization, which modifies pixel values to achieve a uniform
distribution. Holomorphic filtering [219], FFT and high pass filter can be used to
compensate for rough illumination or specular reflection variations in order to obtain the
high contrast lesion images.
For the removal of black frames produced in the digitization process, the author in [220]
[221] proposed an iterative algorithm based on the lightness component of the HSL (Hue-
Saturation-Lightness) color space. In order to remove air bubbles and dermoscopy-gels,
adaptive and recursive weighted median filter developed by the author in [222] that can be
CHAPTER 4 Pre-processing and Segmentation
98
utilized. This type of median filter has an edge determined capability. A method that can
remove bubbles with bright edges was introduced in [223] where the authors utilized a
morphological top-hat operator followed by a radial search procedure. Line detection
procedure based deference of Gaussian on the 2-D results of Gaussian (DOG) [224] and
exemplar-based object removal algorithm[225] can be used for removing dark lines like
ruler marking. In most cases, image smoothing effectively removes the skin lines and
blood vessels.
Most images are affected to some extent by noise that is unexplained variation in data:
disturbances in image intensity which are either uninterpretable or not of interest. Image
analysis is often simplified if this noise can be filtered out[226]. Researchers summarised
the following key points about filters [226]:
• Filters re-evaluate the value of every pixel in an image. For a particular pixel, the new
value is based of pixel values in a local neighbourhood, a window centered on that pixel,
in order to: reduce noise by smoothing, and/or enhance edges.
• Filters provide an aid to visual interpretation of images, and can also be used as a sign
to further digital processing, such as segmentation. Filters are linear or nonlinear,
– Linear filters are well understood and fast to compute, but are incapable of smoothing
without simultaneously blurring edges.
– Nonlinear filters can smooth without blurring edges and can detect edges at all
orientations simultaneously, but have less secure theoretical foundations and can be
slow to compute.
The experiment in this thesis used three filters winner filter, Gabor Filter and adaptive
median filter.
Wiener filtering is a general way of finding the best reconstruction of a noisy signal [227]:
• applies in any orthogonal function basis,
• Different bases give different results,
CHAPTER 4 Pre-processing and Segmentation
99
One way of filtering is to just set components with too much noise to zero. Wiener filtering
is better; it gives the optimal way of narrowing off the noisy components, so as to give the
best reconstruction of the original signal. It can be applied in spatial basis (delta functions,
or pixels), Fourier basis (frequency components), wavelet basis, etc. Different bases are not
equivalent, because, in particular problems, signal and noise distribute differently in them.
A lot of signal processing work is to find the right basis for particular problems in which
signal is most concentrated. The 2-D adaptive noise-removal filtering, wiener2 low pass-
filters a grayscale image that has been corrupted by constant power additive noise.
Wiener2 uses a pixel wise adaptive Wiener method based on statistics estimated from a
local neighborhood of each pixel. Equation 4.3 showed Matlab form [228]:
J = wiener2 (I, [m n], noise) 4.3
Where filters the image using pixel wise adaptive Wiener filtering, using neighbourhoods
of size to estimate the local image and . The
argument, m and n default values is 3 or 5. The additive noise (G Gaussian white
noise) power is assumed to be noise.
Wiener filtering is best when the basis most cleanly separates signal from noise, and that’s
clear in spatial (pixel) basis, when Wiener filter is applied to the difference between an
image and a smoothed image.
Spatial (pixel) basis [227]: It doesn’t make sense to use the pure pixel basis, because there
is no particular separation of signal and noise separately in each pixel. But a closely related
method is to decompose the image into a smoothed background image (x), and then to take
deviations from this as estimating signal power (S) with noise power (N), as showed in
equation 4.4:
= 4.4
‘Hood’ might be a 5x5 neighborhood centered on each point, as showed in equation 4.5 &
4.6:
= 1/ 4.5
= + 4.6
CHAPTER 4 Pre-processing and Segmentation
100
From equation 4.6, you can adjust Wiener parameter part ( ).
One of the recent studies on mathematical modeling of visual cortical cells [229] suggested
a tuned band pass filter bank structure. These filters are found to have Gaussian transfer
functions in the frequency domain. Thus, taking the Invers Fourier Transform of this
transfer function, it can get a filter characteristics closely resembling to the Gabor Filters.
The Gabor filter is basically a Gaussian (with variances and along and -axes
respectively) modulated by a complex sinusoid (with center frequencies and along
and -axes respectively).
Gabor filters are used mostly in shape detection and feature extraction in image processing.
The Gabor Filters have received extensive attention because the characteristics of certain
cells in the visual cortex of some living thing can be approximated by these filters. In
addition these filters have been shown to possess optimal localization properties in both
spatial and frequency domain and thus are well suited for quality segmentation
problems[230] [231]. Gabor filters have been used in many applications, such as quality
segmentation, target detection, fractal dimension management, document analysis, edge
detection, retina identification, image coding and image representation[231]. A Gabor filter
can be viewed as a sinusoidal plane of particular frequency and orientation, modulated by a
Gaussian envelope. It can be written as showed in equation 4.7:
4.7
Where s is a complex sinusoid, known as a carrier, and is a 2-D Gaussian
shaped function, known as envelope. The complex sinusoid is defined as follows in
equation 4.8,
4.8
Where define the special frequency of the sinusoid and .
The 2-D Gaussian function is defined as displayed in equation 4.9,
4. 9
Thus the 2-D Gabor filter can be written as in equation 4.10:
CHAPTER 4 Pre-processing and Segmentation
101
4. 10
The frequency response of the filter is:
4.11
Where , 4.12
This is equivalent to translating the Gaussian function by in the frequency
domain. Thus the Gabor function can be thought of as being a Gaussian function shifted in
frequency to position i.e. at a distance of from the origin . In
the above equations (4.11) & (4.12), ( , ) are referred to as the Gabor filter spatial
central frequency. The parameters are the standard deviation of the Gaussian
envelope along and directions and determine the filter bandwidth.
Gabor Filter Bank Application
Gabor bank of filters used to get feature images. The filters used with two different values
of angle and two different frequencies of cosine function in the bank. It means each image
can be convolved by 4 Gabor filters as described by[232] as shown in figure 4.3.
Figure 4.3. Gabor filter bank for hair detection:
8 filters of 2 frequencies F1 and F2 and for 4 angles A = 0, 45, 90 and 135 grades.
CHAPTER 4 Pre-processing and Segmentation
102
The median filter is normally used to reduce noise in an image, somewhat like the mean
filter. However, it often does a better job than the mean filter of preserving useful detail in
the image. It works like the mean filter; the median filter considers each pixel in the image
in turn and looks at its nearby neighbors to decide whether or not it is representative of its
surroundings. Instead of simply replacing the pixel value with the mean of neighboring
pixel values, it replaces it with the median of those values. The median is calculated by
first sorting all the pixel values from the surrounding neighborhood into numerical order
and then replacing the pixel being considered with the middle pixel value. (If the
neighborhood under consideration contains an even number of pixels, the average of the
two middle pixel values is used.) Refer to the example in figure 4.4.
Figure 4.4, calculating the median value of a pixel neighborhood.
As can be seen, the central pixel value of 150 is rather misleading of the surrounding pixels
and is replaced with the median value: 124. A 3×3 square neighborhood is used here;
larger neighborhoods will produce more severe smoothing. Matlab forms median filter as
displayed in equation 4.13:
B = medfilt2 (A, [m n]) 4.13
Where equation 4.11 performs median filtering of the matrix in two dimensions. Each
output pixel contains the median value in the m-by-n neighborhood around the
corresponding pixel in the input image. pads the image with 0s on the edges, so
the median values for the points within of the edges might appear distorted.
Median filtering is a nonlinear operation often used in image processing to reduce noise. A
median filter is more effective than convolution when the goal is to instantaneously reduce
CHAPTER 4 Pre-processing and Segmentation
103
noise and reserve edges. The argumenta have the default
neighborhood.
A new switch median filter is presented for suppression of impulsive noise in image. The
proposed filter is Modified Adaptive Center Weighted Median (MACWM) filter with an
adjustable central weight obtained by partitioning the observation vector space[233].
The framework of the MACWM filter is illustrated in Figure 4.5. It is composed of four
parts: a median filter, a set of threshold by Fuzzy C-means (FCM), training the center
weight each block by Least Mean Square (LMS) algorithm, and a decision as to whether
noises exist or not.
At first, according to observation vector space (input image); median of input image will
be calculated. We propose FCM algorithm to partition observation vector space to M
block, and related weights to any blocks will be trained by using the LMS algorithm. The
output value y (k) of the MACWM filter (Figure 4.6) at the processed pixel x (k) is
obtained as follows in equation 4.14:
4.14 That the usual output value from a median filter is denoted as m (k) at location k in a filter
window of size 2n + 1 as follows in equation 4.15:
4.15
Where MED means the median operation.
The MACWM filter achieves its effect through the linear combinations of the weighted
output of the median filter and the related weighted input signal. Here, w (k) denotes the
membership function indicating to what extent an impulsive noise is considered to be
located at the pixel x (k). If w (k) = 1, an impulsive noise is considered to be located at
pixel x (k), and the output value of the filter is equal to the output value of the median
filter. If w (k) = 0, an impulsive noise is considered not to be located at pixel x (k), and the
output value is the same as the input x (k); that is, the pixel x (k) is left unchanged. To
judge whether an impulsive noise exists or not at the input pixel x (k), the membership
function w (k) should take a continuous value from 0 to 1. Therefore, the major concern of
CHAPTER 4 Pre-processing and Segmentation
104
the MACWM filter is how to decide the value of the membership function w (k) at the
pixel x (k).
Figure 4.5. The basic structure of the MACWM filter
In general, the amplitudes of most impulses are more prominent than the fine changes of
signals[234]. Thus, the following two variables can be defined to generate the observation
vector[235] . (Partitioning of observation vector space)[233]
Figure 4.6. Shows the output image of Adaptive median Filtering.
4.3. Image preprocessing: enhancement
Image segmentation is one of the fundamental steps in a computer vision and is widely
used in several image processing applications. Segmentation is an important step in the
medical image analyses and classification for radiological evaluation or computer – added
diagnoses. The main purpose of image segmentation is to partition an input image into a
number of meaningful and non-overlapped regions based on some features that help in
distinguishing between different image regions (object and background), by grouping
together neibourhood pixels based on some pre-defined similarity principle. The similarity
standard can be determined using specific properties or features of pixels representing
CHAPTER 4 Pre-processing and Segmentation
105
objects in the image. As well segmentation is a pixel classification technique that allows
the formation of regions of similarities in the image [236].
a class conversion functions.
Medical images occurred in most radiological applications are visually examined by a
physician. The purpose of image enhancement methods is to process a picked image for
better contrast and visibility of features of interest for visual examination as well as
subsequent computer-aided analysis and diagnoses. As described in [236] different medical
imaging modalities provide specific characteristics information about internal organs or
biological skins. Image contrast and visibility of the features of interest depend on the
imaging modality as well as the anatomical regions.
There is no unique general theory or methods for processing all kinds of medical images
for feature enhancement. Specific medical imaging applications such as cardiac,
neurological, breast and genital imaging, present different challenges in image are
processing for feature enhancement. Medical images show characteristic information about
the physiological properties of the structures and skins. But the quality and visibility of
information depends on the imaging modality and the response functions of the imaging
scanner. Medical images from specific modalities need to be processed using a method that
is suitable to enhance the features of interest.
Image-enhancement tasks are usually characterized in two categories [237], [238], [239]:
• Spatial domain methods: These methods employ image pixel values in the spatial
domain based on the distribution statistics of the full image or local regions.
Histogram transformation, spatial filtering, region-growing, morphological image
processing and model-based image estimation methods are some examples in this
category of image and feature enhancement.
• Frequency domain methods: These methods use information in the frequency
domain based on the frequency characteristics of the image. Frequency filtering,
holomorphic filtering and wavelet processing methods are some examples in this
category of frequency representation based image and feature enhancement.
Also model-based techniques such as neural networks and knowledge-based systems are
also used to extract specific features for pattern recognition and classification[240, 241].
CHAPTER 4 Pre-processing and Segmentation
106
The image form itself is referred to as spatial domain, and methods in this category are
based on direct manipulation of pixels in an image. In this part we focus attention of two
important categories of spatial domain processing: 1. intensity (or grey-level)
transformations and 2. Spatial filtering. The approach in our study is referred to as
neibourhood processing, or spatial convolution. Enhancement is highly sensitive and
attracting especially to new researchers in this field. These techniques are general in choice
and have uses in several other branches of digital image processing.
As noted in the spatial domain techniques operate directly on the pixels of an image. The
spatial domain processes are expressed by the following equation 4.16:
4.16
Where is the input image, is the output (processed) image, and is an
operator on , defined over a specified neibourhood about point . In addition, can
operate on a set of images, such as performing the addition of images for noise
reduction. The principle approach for expressing a spatial neibourhood about a point
is to use square or rectangular region cantered at . The centre of the region is
moved from pixel to pixel starting at the top, left corner, and as it moves, it involves
different neighborhoods. Operator is applied at each location to yield the output ,
at that location. Only the pixels in the neighborhood are used to computing a value of
at .
The simplest form of the transformation is dependent only on the intensity values and
not obviously on , intensity transformation function frequently are written in the
simplified form as showed in equation 4.17
4.17
Where denotes the intensity of and the intensity of , both at any corresponding point
in the images.
In general, a histogram is the estimation of the probability distribution of a particular type
of data. An image histogram is a type of histogram which offers a graphical representation
CHAPTER 4 Pre-processing and Segmentation
107
of the tonal obviously distribution of the grey values in a digital image. By viewing the
image’s histogram, it can analyse the frequency of appearance of the different grey levels
contained in the image. In Figure 4.7 we can see an image and its histogram. The
histogram shows that the image contains only a fraction of the total range of grey levels. In
this case there are 256 grey levels and the image only use values between approximately
50-100. Therefore this image has low contrast.
Intensity transformation functions based on information extracted from image intensity
histograms show a basic role in the image processing, in areas such as enhancement,
compression, segmentation, and description. We focus on using histograms for image
enhancement.
(a) Showed: grayscale image (b) Showed: Histogram out.
Figure 4.7 Showed: (a) grayscale image, (b) Histogram out.
The histogram of a digital image with total intensity levels in the range is defined
as the discrete function as showed in 4.18:
( k) = k 4.18 Where is the intensity level in the interval [0, G] and is the number of pixels in
the image whose intensity level is . The value of is 255 for images of class unit 8,
65535 for images of class unit 16, and 1.0 for images of class double. Remember that
indices in Matlab cannot be 0, so corresponds to intensity level 0; corresponds to
intensity , and so on with corresponding to . Note that for
images of class unit8 and unit16. Generally, it is good to work with normalised histograms,
obtained simply by dividing all elements of by the total number of pixels in the
image, which we indicated by equation 4.19:
CHAPTER 4 Pre-processing and Segmentation
108
4.19
For . From basic probability, we recognise P (rk) as an estimate of the
probability of occurrence of intensity level k. The core function in the Matlab toolbox for
dealing with image histograms is which has the following basic syntax, equation
4.20:
4.20
Where is the input image, is its histogram, h (rk) and is the number of bins used in
forming the histogram (if is not including in the argument, b= 256 is used by default). A
bin is simply a subdivision of the intensity scale. For example, if we are working with
unit8 images and we let b=2, then the intensity scale is subdivided into 2 ranges: 0 to 127
and 128 to 255. The resulting histogram will have 2 values: h (1) equal to the number of
pixels in the image with values in the interval [0, 127], and h (2) equal to the number of
pixels with values in the interval [128, 255]. We obtain the normalized histogram simply
by using equation 4.21:
4.21
Where gives the number of the elements in array (i.e. the number of pixels in the
image).
This section describes a method of imaging processing that allows medical images to have
better contrast. This is attained via the histogram of the picture, using a method that allows
the areas with low contrast to gain higher contrast by spreading out the most common
intensity values. When pathologies are present in an image, trying to define the area of the
lesion or object of interest may be a challenge, because different structures are usually
covered one over the other. For example, in the case of the chest the heart, lungs, and
blood vessels are so close together that contrast is critical for achieving an accurate
diagnosis.
As specified earlier, basically the histogram equalization spreads our intensity values along
the total range of values in order to achieve higher contrast. This method is especially
useful when an image is represented by close contrast values, such as images in which both
the background and foreground are bright at the same time, or else both are dark at the
CHAPTER 4 Pre-processing and Segmentation
109
same time. For example, the result of applying histogram equalization to the image is
presented and the image’s contrast has been improved. The original histogram in figure 4.6
has been stretched along the full range of grey values, as can be notice in the equalized
histogram in Figure 4.8.
Figure 4.8. Showed: intensity image using
There are a number of different types of histogram equalization algorithms, such as:
1. Histogram expansion, (Simple and enhance contrasts of an image).
2. Local Area Histogram Equalization (LAHE) (Offers an excellent enhancement of
image contrast).
3. Cumulative histogram equalization (Has good performance in histogram
equalization).
4. Bar sectioning, (Easy to implement) and
5. Odd sectioning (Offers good image contrast).
As previously explained that intensity levels are continuous quantities normalized to the
range and let denote the probability density function of the intensity
levels in a given image, where the subscript is used for differentiating between the
of the input and output images. Suppose that we perform the following information on the
input levels to obtain output (processed) intensity levels, as in equation 4.22:
4.22
Where is a dummy variable of integration. As presented in [237]: that the probability
density function of the output levels is uniform; that is, equation 4.23:
s 4.23
CHAPTER 4 Pre-processing and Segmentation
110
In other words, the preceding transformation makes an image whose intensity levels is
equally likely, and, in addition, covers the entire range The net result of this
intensity-level equalization process is an image with increased dynamite specified range,
which will be likely to have higher contrast. Note that the transformation function is really
nothing more than the cumulative distribution function When dealing with discrete
quantities we work with histograms and call the proceeding technique histogram
equalization, although, in general, the histogram of the proceed image will not be uniform,
due to the separate nature of the variables. Reference to denote the
histogram associated with the intensity levels of a given image, and the values in the
normalised histogram are approximations to the probability of occurrence of each intensity
level in the image. For separate quantities we work with summations, and the equalizations
transformation becomes, as mentioned in equation 4.24:
4.24
, where sk is the intensity value in the output (proceed) image
corresponding to value rk in the input image.
Histogram equalization is implemented in the Matlab toolbox by function which
has the syntax, equation 4.25:
4.25
Where the input image and is the number of intensity levels specified for the output
image. If is equal to (the total number of possible level in the input image), then
implements the transformation function, T (rk), directly. If levy is > L, then
attempts to distribute the levels so that they will approximate a flat histogram.
Unlike , the default value in the is . For the most part, we use the
maximum possible number of levels (generally 256) for because this produces a true
implementation of the histogram – equalisation just described.
CHAPTER 4 Pre-processing and Segmentation
111
As explained above under intensity transformation the neighborhood processing consists
of: defining a Centre point, ; performing an operation that involves only the pixels in
a predefined neighborhood about that Centre point; letting the result of that operation be
the “response” of the process at that point; and repeating the process for every point in the
image. The process of moving the Centre point creates new neighborhoods, one of each
pixel in the input image. The two principle terms used to identify this operation are
neighborhood processing and special filtering, with the second term being more powerful
if the competitions done on the pixels of the neighborhoods are linear, the operation is
called linear spatial filtering (the term spatial convolution also used); otherwise it’s called
nonlinear spatial filtering.
Linear spatial filtering
The concepts of linear filtering have its roots in the use of the Fourier transform for signal
processing in the frequency domain. Use of the term linear spatial filtering separates this
type of process from frequency domain filtering.
Nonlinear spatial filtering
Generally a tool for generating nonlinear special filters in is function , which
generates order-statistic filters (also called rank filters). These are nonlinear spatial filters
whose response is based on ordering (ranking) the pixels contained in an image
neighborhood which are then replacing the value of the Centre pixel in the neighborhood
with the value determined by the ranking results. We focus our attention on nonlinear
filters generated by . The syntax of function is, displayed in equation
4.26:
4.26
The function creates the output image by replacing each element of by the
element in the sorted set of neighbours specified by the non 0 elements in domain. The
domain is a matrix of and that specify the pixel locations in the
neighbourhood that are to be used in the computation. In this sense the domain acts like a
mask. The pixels in the neighborhood that correspond to 0 in the domain matrix are not
used in the computation. For example to implement a min filter (order 1) of size we
use the equation 4.27
CHAPTER 4 Pre-processing and Segmentation
112
4.2
In this formulation the 1 denotes the sample in the ordered set of samples, and
ones creates an matrix of , indicating that all samples in the
neighbourhood are to be used in the computation. In the terminology of statistics, a
(the first sample of an ordered set) is referred to as the percentile
Similarity; the percentile is the last sample in the ordered set, which is
sample. This corresponds to a , which is implemented using the equation 4.28:
4.28
The best known order statistic filter in digital image processing is the median filter, which
corresponds to the 50th percentile. We can use Matlab function median in to
create a median filter equation 4.29:
4.2
Where median simply computes the median of the ordered sequence
Function median has the general equation 4.30:
4.30
Where vector is whose elements are the median of along dimension . Practically,
the Matlab toolbox provides a specialised implementation of the
equation 4.31:
4.31
Where defines a neighborhood of size over which the medium is computed,
and specifies 1 of 3 possible border padding options: (the default)
in which is extended symmetrically by mirror-reflecting it across its
border, and , in which is padded with 1’s if it is of class double and with 0’s
otherwise. The default form of this function is, showed in equation 4.32:
4.32
This uses ( neighbourhoods to compute the medium and cloths the border of the input
with .
CHAPTER 4 Pre-processing and Segmentation
113
We used Adaptive Median Filter, as filtering techniques [242], this filter ignores
edge effects and boundary conditions as such. The output is a collected version of the
original image, as displayed above, where the amount collected is equal to the maximum
window size vertically and horizontally.
All the above mentioned strategies are meant to facilitate the segmentation and feature
extraction stages which consequently lead to better diagnostic results.
4.4. Image Segmentation
Segmentation subdivides an image into its principal regions or objects. The level to which
the subdivision is carried depends on the problem being solved. That is, segmentation
should stop when the objects of interest in an application have been isolated. Segmentation
of significant images is one of the most difficult tasks in image processing. Segmentation
accuracy determines the final success or failure of computerized analysis procedures. For
this reason, considerable care should be taken to improve the probability of rugged
segmentation. Segmentation algorithms for monochrome images generally are based on
one of two basic properties of image intensity values: discontinuity and similarity. In the
first category, the approach is to partition an image based on immediate changes in
intensity, such as edges in an image. The principle approaches in the second category are
based on partitioning an image into regions that are similar according to a set of predefined
criteria. Edge detection in particular has been an essential of segmentation algorithms for
many years. The edge detection is followed by the introduction to thresholding techniques.
Thresholding also is a fundamental approach to segmentation that enjoys a significant
degree of popularity, especially in applications where speed is an important factor.
Image segmentation methods can be broadly classified into three categories:
1. Edge-based methods, in which the edge information is used to determine
boundaries of objects. The boundaries are then analysed and modified, if needed, to
form closed regions belonging to the objects in the image.
2. Pixel-based direct classification methods, in which heuristics or estimation
methods derived from the histogram statistics of the image, are used to form closed
regions belonging to the objects in the image.
CHAPTER 4 Pre-processing and Segmentation
114
3. Region-based methods, in which pixels are analysed directly for a region growing
process based on a pre-defined similarity principle to form closed regions
belonging to the objects in the image.
Once the regions are defined, features can be computed to represent regions for
characterization, analysis and classification. These features may include shape and texture
information of the regions as well as statistical properties, such as variance and mean of
grey values.
Edge-based methods use a spatial filtering method to compute the first-order or second-
order gradient information of the image. Sobel or directional derivative masks can be used
to compute gradient information as showed in figure 4.9. The Laplacian mask can be used
to compute second-order gradient information of the image. For segmentation purposes,
edges need to be linked to form closed regions. Gradient information of the image is used
to track and link relevant edges. In addition, it has to deal with the worries in the gradient
information due to noise and objects in the image.
Figure 4.9. Sobel Edge Detection Image.
Figure 4.10.Binary Output Image Otsu’s Method[243].
The gradient magnitude and directional information from the Sobel horizontal and vertical
direction masks can be obtained by convolving the respective and masks with the
image as , equations 4.33 & 4.34:
CHAPTER 4 Pre-processing and Segmentation
115
4.34
+
Where, equation 4.35 represents the magnitude of the gradient that can be approximated as
the sum of the absolute values of the horizontal and vertical gradient images obtained by
convolving the image with the horizontal and vertical masks and .
Because of its in-built properties and simplicity of implementation, image thresholding
enjoys a central position in applications of image segmentation. We discuss ways of
choosing the threshold value automatically, and we consider a method for changing the
threshold according to the properties of local image neighborhoods.
Suppose that the intensity histogram shown in Figure 4.11 corresponds to an
image composed of light objects on a dark background, in such a way that object
and background pixels have intensity levels grouped into two main modes. One obvious
way to extract the objects from the background is to select a threshold that separates
these modes. Then any point for which is called an object point;
otherwise, the point is called a background point. In other words, the threshold image
is defined as equation 4.36:
correspond to objects, whereas correspond to the
background. When is a constant, this approach is called global thresholding.
As explained above, we can result with:
Histogram modality analysis: in between peaks, as
displayed on figure 4.10.
CHAPTER 4 Pre-processing and Segmentation
116
Maximal separation of classes (Otsu, 1978):
This matter explained in figures 4.11 and 4.12 below:
Figure 4.11, selected thresholds in valleys between peaks
Figure 4.12, obtaining the best possible separation measure for two classes.
Improving the histogram for better peak separation: Use of gradient to improve
histogram by combine intensity and gradient information for better separation of objects
and background, as displayed in figure 4.13.
Figure 4.13, selecting a threshold by visually analyzing a bimodal histogram.
(Principle of histogram peak separation).
One way to choose a threshold is by visual inspection of the image histogram. The
histogram clearly has two separate modes; as a result, it is easy to choose a threshold that
separates them. Another method of choosing is by trial and error, picking different
thresholds until one is found that produces a good result as judged by an observer. This is
particularly effective in a communicating environment, such as one that allows the user to
change the threshold using a wider threshold (graphical control) such as a slider and the
result immediately.
For choosing a threshold automatically [237], describes the following significant
procedure:
CHAPTER 4 Pre-processing and Segmentation
117
1. Select an initial estimate for , (a suggested initial estimate is the mid-point
between the minimum and maximum intensity values in the image).
2. Segment the image using . This will produce two groups of pixels , consisting
of all pixels with , and , consisting of pixels with .
3. Compute the average intensity values and for the pixels in regions and
4. Compute a new threshold value: (equation 4.37)
5. Repeat steps 2 through 4 until the difference in .
There is toolbox in Matlab which provides a function called gray thresh that computes a
threshold using Otsu’s method [243]. To examine the preparation of this histogram-based
method, we start by handling the normalized histogram as an isolated probability density
function, as in equation 4.38:
Where the total number of pixels in the image, is the number of pixels that have
, and is the total number of possible intensity levels in the image. Now suppose
that a threshold is chosen such that is the set of pixels with levels and
is the set of pixels with levels . Otsu’s method chooses the
threshold value that maximizes the between-class variance , which is defined as in
equation 4.39:
Where equations (34) – (39) displayed the component of as follows:
CHAPTER 4 Pre-processing and Segmentation
118
The gray thresh function uses Otsu's method, which chooses the threshold to minimize the infraclass variance of the black and white pixels. This gray thresh function takes an image which is used to compute its histogram, and then finds the threshold value that maximize .The threshold is returned as a normalized value between 0.0 and 1.0. The calling syntax for gray thresh is, as in equation 4.45:
4.45
Where is the input image and computes a global threshold (level) that can be used to
convert intensity image to a binary image with (im2bw), as displayed above in figure 4.9.
is a normalized intensity value that lies in the range [0, 1]. Because the threshold is
normalized to range , it must be scaled to the proper range before it is used. For
example, if is of class unit 8, we multiply by 255 before using it.
To determine an optimal global grey value threshold for image segmentation, parametric
distribution based methods can be applied to the histogram of an image [237] [244]. Let us
assume that the histogram of an image to be segmented has two normal distributions
belonging to two respective classes such as background and object. Thus the histogram can
be represented by a mixture probability density function as in equation 4.46:
4.46
Where p1 (z) and p2 (z) are the Gaussian (normal) distributions of ,
respectively, with the class probabilities of and such that in equation 4.47:
. 4.47
Using a grey value threshold , a pixel in the image can be classified to
in the segmented image as in equation 4.48:
4.48
Let us define the error probabilities of misclassifying a pixel as in equation 4.49 and 4.50:
4.49
And
CHAPTER 4 Pre-processing and Segmentation
119
Where ) equation” (4.49) and ( ) equation (4.50) are, respectively, the
probability of incorrectly classifying a class 1 pixel to pixels to
. The overall probability of error in pixel classification using the threshold is then
expressed as in equation 4.51:
4.51
For image segmentation, the objective is to find an optimal threshold that minimizes the
overall probability of error in pixel classification. The optimization process requires the
parameterization of the probability density distributions and possibility of both classes.
These parameters can be determined from a model or set of training images [245] [244]
[246] [247].
Let us assume and to be the standard deviation and mean of the Gaussian
probability density function of the ( for two classes) such as is explained in
equation 4.52:
4.52
The optimal global threshold can be determined by finding a general solution that
minimizes Equation (4.50) with the mixture distribution in Equation (4.45) and thus
satisfies the following quadratic expression [248] as in equation 4.53:
4.53
Where equations 4.54 – 4.56 display the components displays of equation (4.53)
4.54
4.55
4.56
If the variances of both classes can be assumed to be equal to , the optimal threshold
can be determined as in equation 4.57:
4.57
CHAPTER 4 Pre-processing and Segmentation
120
It should be noted that in the case of equal probability of classes, the above expression for
determining the optimal threshold is simply reduced to the average of the mean values of
the two classes.
In the histogram-based pixel classification method for image segmentation, the grey values
are partitioned into two or more clusters depending on the peaks in the histogram to obtain
thresholds. The basic concept of segmentation by pixel classification can be extended to
clustering the grey values or feature vectors of pixels in the image. In this approach,
multiple parameters of interest are to be segmented or example, a feature vector may
consist of grey value, contrast and local texture measures for each pixel in the image. A
color image may have additional color components in a specific representation such as
Red, Green and Blue components in the colour coordinate system that can be
added to the feature vector. Magnetic Resonance (MR) or multi-modality medical images
may also require segmentation using a multi-dimensional feature space with multiple
parameters of interest.
Images can be segmented by pixel classification through clustering of all features of
interest. The number of clusters in the multi-dimensional feature space thus represents the
number of classes in the image. As the image is classified into cluster classes, segmented
regions are obtained by checking the neighborhood pixels for the same class label.
However, clustering may produce split regions with holes or regions with a single pixel.
After the image data are clustered and pixels are classified, a post-processing algorithm
such as region growing, pixel connectivity or rule-based algorithm is usually applied to
obtain the final segmented regions [249] [250]. There are a number of algorithms
developed for clustering in the literature and used for a wide range of applications [244]
[251] [249] [250] [252].
Clustering is the process of grouping data points with similar feature vectors together in a
single cluster while data points with different feature vectors are placed in different
clusters. Thus the data points that are close to each other in the feature space are clustered
together. The similarity of feature vectors can be represented by a suitable distance
measure such as Euclidean or Mahalanobis disassociated with the distribution of the
corresponding feature vectors of the data points in the cluster. The formation clusters is
CHAPTER 4 Pre-processing and Segmentation
121
optimized with respect to an objective function involving pre-specified distance and
similarity measures along with additional constraints such as smoothness.
This chapter ‘pre-processing and segmentation’; includes introduction explaining extensive
research in investigation of the varied presentations and physical characteristics of
melanoma, the clinical diagnostic accuracy remains sub optimal. Thus, a growing interest
has been developed in the last two decades in the automated analysis of digitized images
obtained by ELM techniques to assist clinicians in differentiating early melanoma from
benign skin lesions; image types. This chapter also covers general approach of developing
a computer aided design (CAD) system; and presented the four stages of CAD system for
skin lesions including all types of filters. The explanation of image segmentations are also
included in all edge detection operations.
CHAPTER 5 Image Representation and Analysis
122
CHAPTER 5
Image Representation and Analysis
5.1. Introduction
Modern biology can benefit from the advancements made in the area of machine learning
[253]Attention should be taken when judging the superiority of some machine learning
approaches over other categories of methods. It is claimed [253] that the success or failure
of machine learning approaches on a given problem is sometimes a matter of the quality
keys used to evaluate the results, and these may vary strongly based on the expertise of the
user. Of special concern with supervised applications is that all steps involved in the
classifier design (selection of input variables, model training, etc.) should be cross-
validated to obtain a balanced estimation for classifier accuracy. For instance, selecting the
features using all available data and subsequently cross-validating the classifier training
will produce a positively biased error estimate. Because of insufficient proof arrangements
many studies published in the literature as successful have been shown to be overoptimistic
[61]. It should be clear from the narrative examples used in this study that choice, tuning,
and diagnosis of machine learning applications are far from automatic.
5.2. Machine Learning
The term machine learning refers to a set of topics dealing with the design and evaluation
of algorithms that facilitate pattern recognition, classification, and calculation, based on
models derived from existing data. Two facets of automation should be acknowledged
when considering machine learning in broad terms. Firstly, it is intended that the
classification and projection tasks can be accomplished by a suitably programmed
computing machine. That is, the product of machine learning is a classifier that can be
practically used on available hardware. Secondly, it is intended that the creation of the
classifier should itself be highly automated and should not involve too much human input.
This second facet is certainly unlimited but the basic objective is that the use of automatic
algorithm construction methods can minimize the possibility that human biases could
affect the selection and performance of the algorithm. The history of relations between
CHAPTER 5 Image Representation and Analysis
123
biology and the field of machine learning is long and complex. An early technique [254]for
machine learning called the perceptron established an effort to model actual neuronal
behavior, and the field of artificial neural network (ANN) design appeared from this effort.
Further artificial neural network architectures such as the adaptive resonance theory (ART)
[255]and neocognitron [256]were moved from the organization of the visual nervous
system. In the best years, the flexibility of machine learning techniques has grown along
with mathematical frameworks for measuring their reliability, and it is natural to promise
that machine learning methods will improve the efficiency of discovery and understanding
in the increasing volume and complexity of biological data. (More detailed information in
appendix C)
5.3. Main Concepts
Two main examples exist in the field of machine learning: supervised and unsupervised
learning. Both have potential applications in biology. In supervised learning, objects in a
given collection are classified using a set of features. The result of the classification
process is a set of rules that propose projects of objects to classes based exclusively on
values of features. The goal in supervised learning is to design a system able to follow
expect the class membership of new objects based on the available features. In
unsupervised learning is intended to show natural groupings in the data. Thus, the two
paradigms may informally be contrasted as follows: in supervised learning, the data come
with class labels, and we learn how to associate labelled data with classes; in unsupervised
learning, all the data are unlabeled, and the learning procedure consists of both defining the
labels and associating objects with them.
We have assumed certain mathematical notations and concepts to support accurate
characterization of both supervised and unsupervised machine learning methods. These to
solve the problems occurred when a task is defined by a series of cases or examples rather
than by predefined rules. Such problems are found in a wide variety of application
domains, ranging from engineering applications in robotics and pattern recognition
(speech, handwriting, face recognition, skin cancer detection), to Internet applications (text
categorization) and medical applications (diagnosis, prognosis, drug discovery). Given a
number of “training” examples (called data facts, samples, patterns or observations)
associated with desired outcomes, the machine learning process consists of finding the
relationship between the patterns and the outcomes using just the training examples. To
CHAPTER 5 Image Representation and Analysis
124
make it concrete, we conceder the following examples: the data points or examples are
clinical observations of patient and the outcome is the health status: healthy or suffering
from cancer. The goal is to expect the unknown outcome for new “test” examples, e.g. the
health status of new patients. The performance on test data is called “generalisation”. To
perform this task, one must build an analytical model or predictor, which is typically a
function with adjustable parameters called a “learning machine”. The training examples are
used to select an optimum set of parameters.
5.4. Intelligent Image Features Extraction
With the size of image databases increasing dramatically, the practicality of such
information is dependent on how well it can be accessed and searched and how well
knowledge can be extracted from it. Our ability to generate data currently far exceeds our
ability to explore, analyze and let alone understand it.
Recently, there has been a growing and improved interest in intelligent image
management systems based on domain knowledge and applications. Although over the last
decade or so some researchers have developed knowledge-based approaches, they rely
almost exclusively on low-level features such as texture, colour and intensity and have
little high-level features interpretation capability. Also, they mainly operate in pixel
domain, such that images (if compressed) need to be decompressed prior to any analysis or
processing. Such an approach can slow down any application to the point of being
unrealistic.
This thesis will create intelligent methods for solving many difficult high-level image
feature extraction and analysis problems, in which both local and global properties as well
as spatial relations must be taken into consideration. All the techniques will be
implemented to work in pixel and compressed domain (avoiding the inverse transform in
decompression), thus speeding up the whole process. This research will lead to a better
understanding of images and development of efficient methods for other image processing,
pattern recognition and computer vision problems. The research results will have a wide
range of useful applications, including but not restricted to human face image recognition,
medical and biological image analysis, image and video compression, and content-based
image indexing and retrieval.
CHAPTER 5 Image Representation and Analysis
125
Initial investigation of intelligent image analysis and feature extraction techniques will be
done, combining the principles from fields of multiresolution wavelet analysis,
compression, computer vision, texture analysis and information retrieval. Next, by
applying and extending those techniques we aim to solve a class of difficult image feature
extraction problems. By combining various image analysis and signal processing
techniques by develop some new high-level feature extraction methods, thus improving
current state-of-the-art retrieval and classification methods. Using the resulting extracted
features as a first step and input to data mining systems would lead to supreme knowledge
discovery systems [257].
Texture is one of the important features used in image retrieval systems. The methods of
describing texture go down into two major categories: Statistical and Structural. The
primary goal is to determine which texture feature or combination of texture features is
most efficient in representing the spatial distribution of images. Researchers involved in
analyses and comparison of texture features [258], they evaluate both Statistical and
Structural texture features and presented comparison of individual texture features and
combined texture features. The first-order statistics, second-order statistics, Gabor
transform and 2D Wavelet transforms were considered for retrieval. The retrieval
efficiency of the texture features was investigated by means of relevance. According to the
results they obtained[258], it displays that it is difficult to claim any individual feature is
superior to others. The performance depends on the spatial distribution of images. The test
results indicated that Grey Level Co-occurrence Matrix performs well compared to other
features when images are similar. In most of the image categories, Autocorrelation feature
also shows similar performance[258]. It is also noted that the structural texture features are
more effective than the statistical texture features. In case of combination features,
combinations recorded better retrieval rate compared to the performances of those
individual texture features.
The texture defined as the structural pattern of surfaces which is homogeneous in spite of
fluctuations in brightness and color. Most important visual key in identifying regions.
CHAPTER 5 Image Representation and Analysis
126
Main categories divided to: regularity, randomness and directionality[259] as displayed in
figure 5.1.
Figure 5.1 displayed the main categories: texture elements, regularity, randomness, directionality and regularity.
With the current enormous scientific image databases present, it is simple to expect any
human to be able to analyses understand and extract knowledge from it. Fortunately, in
spite of how logically difficult the task of knowledge discovery from such data may
appear, it turns out that the process consists of some computational parts that may be
automated and actually better executed by computers than by humans. This research focus
on intelligent image features extraction, typically by the use of categorized or
multiresolution analysis. By trying many experiments, and the aim of this research is to
improve some algorithms and develop new ones to be able to generalize to various
scientific image analysis problems[260]. Wavelets and other image analysis techniques
have shown themselves useful in a wide array of domains such as classification of tissues
in computed tomography image retrieval in medical databases.
The texture features divided to [257].
• Statistical approach
– From first order statistics (histogram)
– From second order statistics (grey-level co-occurrence matrix)
• Spectral approach -Fourier power spectrum
• Gabor filter
• Multiresolution wavelet
CHAPTER 5 Image Representation and Analysis
127
The grey-level co-occurrence matrix is defined by specifying a displacement vector
and counting all pairs of pixels separated by having grey levels as
displayed on figure 5.2.
Figure 5.2, showed, (a) images (5x 5) matrix with 3 grey levels 0, 1, and 2. (b) The co-
occurrence matrix for d= (1, 1).
The co-occurrence matrices can be computed for 0, 45, 90 and 135 degrees and at
distances 1, 2, 3..., statistical features are computed from the co-occurrence Matrices,
Entropy is a feature which measures the randomness of gay-level distribution. The angular
second-moment is a measure of similarity of the image, the contrast is a measure of the
amount of local variations present in the image, and the correlation is a measure of
linearity in the image.
The texture matrix used was the grey-level co-occurrence matrix (GLCM). In designing
the GLCM for texture representation, there are three fundamental parameters that must be
defined: 1. quantization levels of the image, 2. the displacement, and 3. orientation values
of the measurements.
The definition of GLCM’s [261] is as follows: Suppose an image to be analyses is
rectangular and has columns and rows. Suppose that the grey level appearing at each
pixel is quantized to levels. The texture-context information is specified by the matrix of
relative frequencies with two neighbouring pixels separated by distance occur on the
image; one with grey level and the other with grey level . Such matrices of grey-level co-
occurrence frequencies are a function of the angular relationship and distance between the
neighbouring pixels. Let be the th entry in a normalized GLCM. The mean
( ) and standard deviations ( ) for the rows and columns of the matrix are in
equations (5.1) – (5.4):
5.1
CHAPTER 5 Image Representation and Analysis
128
5.2
5.3
5.4
As per haralick et el [261] suggests that a set of 28 textural features which can be extracted
from each of the gray-tone spatial-dependence matrices. The following 5.5 equations
define these features.
Notation
(i) ith entry in the marginal-probability matrix obtained by summing the rows of
, = . 5.5
where Number of distinct gray levels in the quantized image equations 5.6, 5.7,
and 555.8:
respectively.
5.6
5.7
5.8
The Gabor Filters have received considerable attention because the characteristics of
certain cells in the visual cortex (the outer layer of an internal organ) of some creatures can
be approximated by these filters. In addition these filters have been shown to keep optimal
localization properties in both spatial and frequency domain and thus are well suited for
texture segmentation problems. Gabor filters have been used in many applications, such as
texture segmentation, target detection, fractal dimension management, document analysis,
edge detection, retina identification, and image coding and representation. A Gabor filter
can be viewed as a sinusoidal plane of particular frequency and orientation, modulated by a
Gaussian envelope. It can be written as displayed on equation 5.9:
= 5.9
CHAPTER 5 Image Representation and Analysis
129
Where is a complex sinusoid, known as a carrier, and is a
shaped function, known as envelope. The complex sinusoid is defined as follows, equation
5.10:
5.10
In this thesis an investigation of the validity of Gabor filter banks for feature extraction in
an image recognition context is presented [259]. in equation 5.11:
= 5.11
It showed a response from convolving image sample with filter, as displayed in equation
5.12, and figure 5.3,
5.12
Figure 5.3, displayed the response from convolving image sample with filter.
Discrete wavelet transform (DWT) can be used to decomposing the image into non-
overlapping sub bands as shown in Figure 5.4, a) [262], the mean and standard deviation of
the decomposed image portions are calculated. These mean and standard deviation are
treated as the texture features for image comparison.
The block diagram in figure 5.4, b) showed the methodology of the – Gabor
approach as follows. Firstly, The is applied to the biomedical image to obtain its high
frequency image component contains most of the desired information about the biological
tissue[26]. Secondly, a bank of Gabor filters with different scales and orientations is
CHAPTER 5 Image Representation and Analysis
130
applied to the high-frequency image to obtain Gabor-filtered images along different spatial
orientations. Thirdly, statistical features are extracted from the Gabor filtered images.
Finally, the is used to classify the resulting feature vector for final diagnosis.
(a)
(b)
Figure 5.4, (a) Wavelet decomposition of an image (b) Block diagram of the decomposition of an image
Figure 5.5 displayed a block diagram of the discrete wavelet transforms (DWT – Gabor approach).
The two-dimensional discrete wavelet transform (2D-DWT) [263] [264] performs a Sub
band coding of an image in terms of spectral spatial/ frequency components, using an
iterative and recursive process. Figure 5.5, illustrates the case of two-level decomposition.
The image is first represented by LH, HL, and HH sub bands that encode the image details
in three directions and an LL sub band which provides an approximation of it. The
approximation images can be decomposed again to obtain second-level detail and
CHAPTER 5 Image Representation and Analysis
131
approximation images, and the process can be repeated for finer analysis as each iterations
doubles the image scale as shown in figure 5.4
This thesis used in total 283 features, six features extracted from first order statistics
(histogram), twenty one features extracted from second order statistics (grey-level co-
occurrence matrix) for each channel and 255 features extracted from Wavelet Packet
Transform (WPT).
The multi-resolution wavelet features produce:
• A 2-level wavelet transform of the image is constructed using the orthonormal
Daubechies filter. (Figure 5.5).
• The following features 1-3, represented by equations 5.13, 5.14 and, 5.15:
1. Wavelet energy = 5.13
2. Variance = 5.14
3. Residual energy = = 5.15
It represented by equation 5.31, = 5.31
Where: is the feature vector for texture class I, is the
feature vector for texture class j, is the number of features, and is the std.
deviation of the feature over the texture classes.
5.5. Feature Extraction and Representation
A data representation (image) must be chosen [236]. Data are represented by a fixed
number of features which can be binary, definite or repeats. A feature is identical to input
variable or quality. It is sometimes necessary to distinguish between “raw” input variables
and “features” that are variables built for the original input variables[265]. In our medical
diagnosis example, the features may be indications that are a set of variables categorizing
the health status of a patient.
CHAPTER 5 Image Representation and Analysis
132
Human expertise, which is often required to convert “raw” data into a set of useful
features, can be complemented by automatic feature construction methods. Feature
construction is a pre-processing stage. Pre-processing transformations may include[265]:
-Standardization: Features can have different scales although they refer to similar objects.
-Normalization: This interprets into the formula: .
-Signal enhancement: The signal-to-noise ratio may be improved by applying signal or
image-processing filters. These operations include baseline or background removal, de-
noising, smoothing, or sharpening. The Fourier transform and wavelet transforms are
popular methods. We refer to preliminary books in digital signal processing [266],
wavelets[267], image processing [268], and morphological image analysis [269].
-Extraction of local features: These techniques convert problem specific knowledge into
the features. It is worth mentioning that they can bring significant improvement.
-Linear and non-linear space embedding methods: When the dimensionality of the data is
very high, some techniques might be used to implant the data into a lower dimensional
space while retaining as much information as possible. Principal Component Analysis
( ) and Multi-Dimensional Scaling ( )) [270].
-Feature discretization: Some algorithms do no handle well continuous data. it may
simplify the data description and improve data understanding[271].
Texture is one of the important characteristics used in identifying objects or regions of
interest in an image, whether the image is a photomicrograph, or a satellite image.
The literature topics in this study are shown in a thematic way rather than in a
methodological way. In this section, we present a unified view of feature selection that
exceeds the filter/wrapper and is inspired by the views of [271]. There are (4) four aspects
of feature extraction:
1. feature construction;
2. feature subset generation (or search strategy);
3. Evaluation criterion definition (e.g. relevance index or predictive power);
CHAPTER 5 Image Representation and Analysis
133
4. Evaluation criterion estimation (or assessment method).
The last three aspects are relevant to feature selection and are schematically
summarized in Figure 5.6. Filters (a) and wrappers (b) differ mostly by the evaluation
Figure 5.6, the three principal approaches of feature Mammographic e selection. The shades show the components used by the three approaches: filters, wrappers and embedded
methods.
criterion. It is usually understood that: filters use criteria not involving any learning
machine, e.g. a relevance index based on correlation coefficients or test statistics, whereas
wrappers use the performance of a learning machine trained using a given feature subset.
5.6. Classification and Feature selection
Researcher on [265] reported that: the problem of feature extraction rotting into (2) two
steps: 1) feature construction: is one of the key steps in the data analysis process, it
CHAPTER 5 Image Representation and Analysis
134
compare the performances obtained with either image. 2) Feature selection: is primarily
performed to select relevant and useful features. It is an important problem for pattern
classification systems. We study how to select good features according to the maximal
statistical need standard based on common information. We first derive an equivalent form,
called minimal-redundancy-maximal-relevance criterion , for first-order
incremental feature selection. Then, we present a two-stage feature selection algorithm by
combining mRMR and other more sophisticated feature selectors (e.g., wrappers). This
allows us to select a compact set of superior features at very low cost. [Superior features =
mRMR + wrappers = very low cost].
Although Support Vector Machine ( ) as classifier directly determine the decision
boundaries in the training step and the method can also provide good generalization in
high-dimensional input spaces, many researchers reported that feature selection is
important for [272]. In general, the major aims of feature selection for classification
are finding a subset of variables that result in more accurate classifiers and constructing
more compact models. An amount of variables can be redundant to the classification
problem and feature selection can filter out the variables that are unrelated for the model.
Three major types of feature selection methods are distinguished, namely filter, wrapper
and embedded methods[28] [273, 274] [275].
• Filter methods
Filter methods employ an initial analysis as a pre-processing step, where a set of features is
selected as input for the classifier using a training data set[276]. Researcher in[277],
proposed that the selection procedure should take the learning algorithm into account.
• Wrapper methods
In wrapper models [276], a subset of variables is selected based on a search by actually
using the classifier. The best subset of features is obtained by estimating and comparing
the performances of the various subsets of features, which are fed to the classifier. An
complete search through the input space is often not feasible, therefore experimental search
methods using backward, forward or stepwise variable selection are employed [278]. In
addition, more sophisticated methods like best-first search are also able to navigate the
space of subsets [279] [276]. For the evaluation of each subset, n-fold cross-validation or
leave-one-out (LOO) cross-validation can be used. In general the wrapper models result in
an increased accuracy because potential effects of the classification algorithm are taken
CHAPTER 5 Image Representation and Analysis
135
into account [276]. On the other hand, in principle, wrapper methods involve the higher
computational cost of a search. A popular algorithm that combines SVMs with a backward
search is that of recursive feature elimination (RFE) [280]. In this approach the role of a
feature is estimated by the change in the cost function when such a feature is removed. For
this purpose, the weights in the SVM model are used while support vectors are fixed for
fast feature elimination. An efficient algorithm for variable selection in high-dimensional
scenarios was proposed in reference[281].
• Embedded methods
In contrast to filter and wrapper methods, embedded methods aim to immediately integrate
the variable selection or weighting procedure in the learning algorithm. Gradient-based
algorithms are used to perform feature selection by minimization of generalization error
bounds. Reference [282] showed the radius-margin bound for is optimized with
respect to feature scaling factors, which reflect the contribution of features to the prediction
rule. The work by[283]. provides derivations for several generalization bounds in
terms of the feature scaling factors[283]. Such formulations can serve as a methodology for
multiple hyper-parameter tuning. Within this category automatic relevance determination
(ARD), aims to identify informative input features as a natural result of optimizing the
model selection criterion [283] [284]. Tuning parameters of and Kernel and other
kernel-based models usually depend on several parameters controlling the model
complexity. In parameters, and the bias term , are determined by solving the
rounded optimization problem, while the positive penalization constant ( ) and the kernel
parameters remain. So as decreases, the width of the margin increases. In the setting of
with or in polynomial kernel, determine the shape of the
decision boundary as well as the dimension and complexity of the matching induced
feature space. A typical a priori setting for the RBF kernel parameter is that of using the
distance between the closest points from different classes. However, applications rely on
the use of a validation set or cross-validation to choose hyper-parameter values. In
practice, cross-validation-based techniques are preferred over generalization error bounds.
Researcher in[285] developed an interesting benchmark study which presents an efficient
methodology for hyper- parameter tuning and model building using Least squares support
vector machines ( ) as benchmark. The approach is based on the closed form
leave-one-out ( ) computation for , only requiring the same computational
CHAPTER 5 Image Representation and Analysis
136
cost of one single training. Performance compares favorably to bound-based
techniques with the advantage of exact and efficient computation.
In many binary classification problems, data sets are often skewed in favor of one class
such that the contributions of false negatives and false positives to the performance
assessment criterion are not balanced. In such cases the sizes of positive and negative
points in the training data are not truly typical for the operational class frequencies. As
mentioned in literature, appropriate prior class probabilities need to be specified for
Bayesian methods and can therefore take an unbalance in the data into account. In case of
standard or a constant bias term correction can be performed[286].
Furthermore, different weights can be introduced for positive and negative data points,
thereby balancing their contributions to the regularized loss function. Taking into account
the unbalance in the data set, the objective functions for SVM and LS-SVM displayed on
equations 5.34 5.35 respectively.
= w + C 5.34
= w + 5.35
With: = ,
Where and , represent the number of positive and negative data points, respectively.
Asymptotically this weighting is equivalent to resampling of the data such that there are an
equal number of positive and negative data points.
Sequential Feature Selection (SFS) for Classification Sequential Feature Selection (SFS)
will insert Reinforcement Learning (RL) into classification tasks, with the objective to
reduce data consumption and associated costs of features during classification [287].
Feature selection with RL explained in [287]. Researcher in [288]reported that the decision
process is not dependent on the internal state of the classifier, which brings their method
closer to straight FS. Researchers in [289] reported that they approach reduces the number
of necessary features to access to a fraction of the full input, down to 12% with RNN
CHAPTER 5 Image Representation and Analysis
137
classifiers. They also demonstrated that SFS is able to deal with weighted feature costs, a
property that exists in plenty of real-world applications.
5.7. SVM for Classification and Regression
Support vector machines (SVM’s) are currently a hot topic in machine learning community
and are becoming popular in a wide variety of biological applications. A support vector
machine is a computer algorithm that learns by example to assign labels to objects [290].
Researchers in [291] presented the advantage of SVM for its generalization capability.
SVM utilizes procedure for reducing the training data points which participates in defining
discriminant function. Thus, the participating data points, which are called support vectors,
are minimized. Data points which do not participate in defining a classifier are ignored.
This can reduce noise and, consequently, can improve the generalization capability. SVM
has proved mostly good performance for classification in varied applications, both for real
and artificial standard benchmarking data [292], including applications in medical fields.
The problem of experiential data modeling is relevant to many engineering applications.
SVM is a classifier which works through deciding an optimal hyperplane which optimally
hyperplane which optimally separates two class data. The optimal hyperplane is also called
a decision surface which is the surface optimally separating two groups of data points.
Suppose there are linearly-separable training-data and the associated
is a class label is for the class and is for the class –1.
For example, the class is for the data which belongs to benign, and –1 is for melanoma.
In the case of two–dimensional space, as shown in Figure 5.7 two groups of data (circles
and squares) can be separated by different hyperplanes. Two possible hyperplanes are
shown in Figure 5.8 (a) and (b). In SVM, the optimal hyperplane is right in the middle of
the two class boundaries and the distance of the two boundaries is maximized. The optimal
hyperplane promises good classification while facing unseen data and provides a good
generalization. In Figure 5.7 the two parallel lines in the left and right of the hyperplane are
supporting hyperplanes. The supporting hyperplanes are parallel to the hyperplane. The
decision surface (or hyperplane) and the supporting hyperplanes are defined as follow:
The decision surface: 5.36
CHAPTER 5 Image Representation and Analysis
138
Supporting hyperplane for class +1: 5.37
Supporting hyperplane for class -1: 5.38
Where the perpendicular is distance from the hyperplane to the origin and is called
bias.
The distance between the two supporting hyperplanes is referred to as a margin . The
optimization problem of is to find the decision boundaries so that the distance
between the supporting hyperplanes is as far as possible. Therefore, the optimization is to
maximize the margin and to keep the decision surface equidistant from the two supporting
hyperplanes. As in Figure 5.8, the margin is constrained by the points which are indicated
by the filled circles and the filled squares. These points are called support vectors.
Because the optimization problem of is to find the optimal hyperplane, the feasible
solutions to the optimization problem are all possible hyperplanes, with their associated
supporting hyperplanes. The constraints are the positions of the supporting hyperplanes in
order not to cross their respective class boundaries. Considering Figure 5.8, the margin
between the two supporting hyperplanes can be defined as (more detailed information
in appendix D)
Figure 5.7: Two-out-of-many separating lines; (a) with smaller margin and (b) with larger margin
5.39
Thus, the maximum margin is obtained by maximizing Eq. 5.40, as
5.40
CHAPTER 5 Image Representation and Analysis
139
5.41
Eq. 5.41 can also be stated as:
5.42
Because optimizing over is the same as optimizing . Eq. 5.42 can be formulated as:
5.43
As mentioned above, the constraints of the optimization problem are the positions of the
supporting hyperplanes which do not cross their respective class boundaries.
Mathematically, it can be formulated as:
+b for all ( ) such that 5.44
+b for all ( ) such that 5.45
Using a more compact formula, it can be stated as:
5.46
Figure 5.8: Margin m between two supporting hyperplanes.
The optimization problem in Equation 5.43 is solved using the Lagrangian. The
construction of the Lagrangian for the optimization problem is:
5.47
where is a Lagrangian multiplier. This gives the Lagrangian optimization problem as:
5.48
Subject to
CHAPTER 5 Image Representation and Analysis
140
5.49
This provides the Lagrangian dual optimization problem as follow:
- 5.50
and the optimal decision surface is defined as (see Appendix E)
5.51
5.52
The SVM which is mentioned above is a linear SVM. The linear SVM needs to be
extended to tackle non-linearly separable data, as data in the real world are, mostly, not
linearly separable. The idea of the extension is by using the kernel trick. Using a kernel
function, data in the input space is transformed to a higher dimensional space, which is
called a feature space. In this feature space it is possible to separate the data using a
hyperplane. The transformation is illustrated in Figure 5.10. The left side of figure 5.10,
the data points cannot be classified by a linear function. After the data points are
transformed in three dimensional spaces, a linear function can be used to classify the data
points. The right side of Figure 5.9, a plane can separate the data points; two triangles are
above the plane and two squares are below the plane.
Figure 5.9: Illustration of mapping using a transform .
Using the kernel trick, the Lagrangian optimization problem is modified by replacing
using a mapping function (x) as:
. 5.53
In the compact form Eq. 5.53 can be written as:
CHAPTER 5 Image Representation and Analysis
141
( . 5.54
where k ( . ) is a kernel function. Kernel functions include:
Radial basis function (RBF), 5.55
Sigmoid, = 5.56
Polynomial, 5.57
Linear, .k ( ) = 5.58
A soft-margin classifier of is useful to tackle imperfect data containing a noisy data
point, such as that which is due to improper measurement. In this soft-margin , a slack
variable is introduced. A slack variable measures how much of an error is introduced by
allowing the supporting hyperplane to be unconstrained by a data point (Figure 5.10). By
introducing a nonnegative slack variable, the corresponding modified constraint is:
5.59
and the optimization problem in 5.43 becomes
5.60
The introduction of the slack variable as in the form of is to prevent the
Construction of trivial solutions; where all training points are considered as noise during
the optimization. is a constant, called the cost, controlling the trade-off between margin
size and error. By this slack variable introduction, the Lagrangian construction of the
optimization problem is (see Appendix F)
5.61
This Lagrangian dual for soft-margin SVM is the same as the Lagrangian dual for hard-
margin . Therefore the objective functions of the optimization problems for both
are the same. The difference is in the constraint. The constraint in the soft-margin
can be stated as:
5.62
CHAPTER 5 Image Representation and Analysis
142
In a case of an imbalanced data point number between two classes, different error weights,
and , are necessary to penalize more heavily the class with the smaller
population[293].
Figure 5.10: Introducing slack variable in soft-margin SVM
5.63
In summary, the decision function of support vector machine can be represented as follow
5.64
In which is the solution of maximizing the following Lagrangian
5.65
Subject to constraints
5.66
5.67
The optimization of Eq. 5.65 uses the method based on sequential minimal optimization
( )[294, 295],[296]. In the optimization, SMO uses analytical steps, instead of
numerical quadratic programming. In each iteration SMO chooses two Lagrangian
multipliers and optimize these two multipliers. Two Lagrangian multipliers at every step is
the smallest number in the optimization because the Lagrangian has to obey a linear
equality constraint (Eq. 5.66).
The optimization which uses an SMO could be faster than the optimization which uses
quadratic programming (QP). Using QP, the optimization for Eq. 5.65 involves a matrix
with a number of elements that equals to the square of the number of training examples.
CHAPTER 5 Image Representation and Analysis
143
This number obviously needs a large memory and can result in a very slow optimization.
The description of the optimization using SMO is presented in Appendix G [297].
SVMs can also be applied to regression problems by the introduction of an alternative loss
function [298]. The loss function must be modified to include a distance measure. Figure
5.12 illustrates four possible loss functions.
The loss function in Figure 5.11(a) corresponds to the conventional least squares error
criterion. The loss function in Figure 5.11(b) is a Laplacian loss function that is less
sensitive to outliers than the quadratic loss function. Huber proposed the loss function in
Figure 5.11 (c) as a robust
Loss function that has optimal properties when the underlying distribution of the data is
unknown. These three loss functions will produce no sparseness in the support vectors.
Researcher in [299] proposed the loss function in Figure 5.11(d) as an approximation to
Huber’s loss function that enables a sparse set of support vectors to be obtained. The
Support Vector Regression divided to: Linear Regression, and Non Linear Regression.
Larger and more complex classification problems have been attacked with .
Researcher in [300] has applied to the exacting problem of face recognition, with
encouraging results. In conclusion, provides a robust method for pattern classification
by minimising over fitting problems by adopting the principle. Use of a kernel
function enables the curse of dimensionality to be addressed, and the solution implicitly
contains support vectors that provide a description of the significant data for classification
CHAPTER 5 Image Representation and Analysis
144
(a) Quadratic (b) Laplace
(c) Huber (d) –insensitive
Figure 5.11: Loss Functions
5.8. Swarm based Support Vector Machine for Melanoma Detection
A hybrid technique of swarm-based support vector machine (SVM) is introduced for
melanoma detection using the pathology images as inputs. In this technique, a particle
swarm optimization (PSO) is proposed to optimize the SVM to detect melanoma.
The structure of the swarm based support vector machine ( ) is presented in Figure
5.12. The input was pathology image, and the output was the melanoma or benign. The
parameters of SVM were optimized using (the optimal parameters of were
found using ). The parameters are . The parameter of deals with
a trade-off between maximum margin and the classification error of with
high provides perfect classification in the training phase which uses a training data set,
but it may be poor in generalization (or bad performance in classification using a test data
set). On the other hand, classification with low might give a bad performance in the
training phase. Therefore, optimal has to be chosen in order to find the optimal
CHAPTER 5 Image Representation and Analysis
145
classification. Likewise, the optimal values of parameters and also need to be found to
obtain the optimal performance.
Figure 5.12: Melanoma detection using swarm based support vector machine.
This research utilizes the algorithm to find the optimal parameters according to
a defined fitness function. Four parameters are optimized; they are (for RBF and
sigmoid kernel functions) or (for polynomial kernel function), and (weight factors
for Benign and melanoma class data, respectively).
The optimization of parameters using can be described as in Figure 5.13.
Suppose a swarm at -th iteration is . It contains particles with dimensions. More
clearly, in this work four parameters are used, and therefore, the is 4. The number of
particles used in the optimization of this work is 50. Each element of the swarm is
presented by , in which and ; is the dimension of a
particle and denotes the number of particles in the swarm. The particles can be
expressed as in the matrix in Figure 5.14.
CHAPTER 5 Image Representation and Analysis
146
Figure 5.13: The pseudo of the PSO for the SVM parameter optimization
The algorithm was initialized by generating random numbers for position and velocity. At
iteration , the position is determined as mentioned by [301]equation 5.68:
5.68
in which the velocity is defined as[301] in equation 5.69:
5.69
is the personal best position, which is the best position of the particle at iteration
is the global best position which is the best position among all the particles from the
first iteration until iteration is an inertia weight factor, which controls the convergence
of the optimization behaviour. and are the acceleration constants, which control the
distance, moved by a particle. and are the uniform random numbers between 0 to
1[302].The constants and are set to 2; and and are the random number in the
range [0, 1].
CHAPTER 5 Image Representation and Analysis
147
Figure 5.14: The particles of the PSO
The velocity is limited in a certain region, which is between and Thus, if
is more than , v is set to and if is less than , v is set to . This
strategy is called velocity clamping, which is used to limit the swarm particles to a search
space[303]. For this optimization, is set to 0.2. Furthermore, the inertia weight is set
as in equation 5.70:
- 5.70
where and are upper and lower inertia weights, respectively, and are set to
1.2 and 0.1, respectively, is the total iteration number and is the iteration number.
The objective of the PSO optimization was to maximize the performance of the
hypoglycaemia detection. The performance was measured in terms of sensitivity and
specificity , as defined in Equations 5.45 and 5.46. Sensitivity was defined as the ratio of
correct detection of melanoma to the actual number of melanoma cases. Specificity was
defined as the ratio of correct detection of benign to the actual number of benign cases.
Thus, the optimization was essentially used to maximize sensitivity and
specificity , and then the fitness function was defined as in equation 5.71:
CHAPTER 5 Image Representation and Analysis
148
5.71
and were, respectively, the sensitivity and specificity of the training and ,
respectively, were the sensitivity and specificity of the validation. The training data set was
used for them training to find a model. The validation data set was used to test
the SVM model during the optimization.
The inclusion of and in the fitness function is to reduce the risk of overtraining[302].
Overtraining could yield a high performance in the training stage, but it might provide a
low performance in the testing. In the fitness function, was set to 0.58 to obtain a higher
sensitivity than specificity. This strategy is to prevent the risk of the low sensitivity of the
melanoma detection. Furthermore, to force the detection to have high sensitivity, parameter
is given by using the definition in equation 5.72:
5.72
By the definition of in Eq. 5.72 the detection is forced to find sensitivity and specificity
of more than 70% and 40%, respectively. Considering Eq. 5.71 and Eq. 5.72, if the
detection provides sensitivity and specificity of more than 70% and 40%, respectively, the
fitness function has the value of more than 10. Conversely, if the detection provides
sensitivity and specificity of less than 70% and 40%, respectively, the fitness function has
the value of less than 2; in this formula the maximum sensitivity and specificity is 1 (or
100%).
This chapter ‘image representation and analysis’ includes introduction; machine learning;
main concepts; intelligent image features extraction such as Texture feature, Grey-Level
Co-occurrence Matrix, Gabor filter feature extraction, Wavelet decomposition of an image;
Discrete Wavelet Transform; Classifier distance metric equation. Feature extraction and
representation includes Methodology that is used in feature extraction; classification and
feature selection. Other machine learning techniques for classification and regression like
Linear support vector machine, Nonlinear support vector machine, Support Vector
Regression; swarm based support vector machine for melanoma detection are also
CHAPTER 5 Image Representation and Analysis
149
discussed for the development of a Melanoma Detection based on the Swarm based
Support Vector machine. Optimization of SVM parameters using PSO and Fitness function
for the optimization are two other areas covered into this chapter.
CHAPTER 6 Experimental Results and Discussion
150
CHAPTER 6
Experimental Results and Discussion
6.1. Introduction
The rate of skin cancer is rapidly increasing throughout the world including Australia [1].
Australia has one of the highest skin cancer rates in the world at nearly four times the rates
in Canada, the US and the UK [76] [304]. It has been estimated 115,000 new cases of
cancer are diagnosed and more than 43,000 people are expected to die from cancer this
year according to the Cancer council of Australian 2010 [305], The Chair of Public Health
Committee pronounced that, more than 430,000 cases treated for non-melanoma, and more
than 10,300 people are treated for melanoma, with 1430 people dying each year [305]. At
least two in three Australians will be diagnosed with skin cancer before 70 years of age
[306], Skin cancers accounted for over 80% of all cancers diagnosed in Australia
[307]each year. Skin cancer costs the health system around $300 million Australian dollar
annually, the highest cost of all cancers. Melanoma has near 95% cure rate if diagnosed
(detected) and treated in the early stages [305]. Visual diagnosing guide to nick-eye error
since it is visually hard for medical professionals to differentiate normal from abnormal
moles, and general practitioners (GP) do not have the core expertise to diagnose skin
cancers. Dermatologist’s data indicate that even in specialized centers diagnostic accuracy
is only about 60% [308], but they are overloaded by GP referrals. In order to reduce the
occurrence of skin cancer, computer- aided detection and diagnosis technologies have been
developed. The extraction of the skin cancer border is a key technology in computer- aided
detection technologies for skin cancer diagnosis. Previously several technologies / methods
have been proposed. Several published classification systems show accuracy rates ranging
from 60%- 92% [309] which coincides with the estimated rates obtained by general
practitioners. These techniques were aiming to be able to provide recommendation for
nonspecialized users. However the variations of diagnosis are sufficiency large and there is
a lack of detail of the test methods. One commercial product, Solar Scan by Polar technics,
has an accuracy rate of 92%. Solar Scan is a complex system, taking high quality
Epiluminescence Light Microscopy (ELM) images and using advanced image analysis
CHAPTER 6 Experimental Results and Discussion
151
techniques to extract a number of features for classification which it makes it unsuitable for
a normal person to use. The traditional imaging is just a recording of what the human eye
can see by digital camera while the Dermoscopy known as Epiluminescence Light
Microscopy (ELM) images needs an experienced professional to get the required image.
This chapter is organized in 6 Experiment Based on Digital and pathological images,
Experiment 1: These experiments investigate skin cancer segmentation using an active
contour model, or snake model, driven by Gradient Vector Flow (GVF). Experiment 2:
These experiments propose two segmentation methods to identify the normal skin cancer
oriented texture and animation edges and then finding the accuracy by using the three
layers back-propagation neural network classifier (BNN). Experiment 3: This experiment
proposes an automated non-invasive system for skin cancer (melanoma) detection based on
Support Vector Machine classification. The proposed system uses a number of features
extracted from the Wavelet or the Curvelet decomposition of the grayscale skin lesion
images from the original color images. Experiment 4: This experiment used a novel
methodology for automatic feature extraction from histo-pathological images and
subsequent classification. These techniques added Gabor filter bank to winner and adaptive
median filters to improve diagnostic accuracy. Then Histogram equalization was used to
enhance the contrast of the images. The extracted features were reduced by using
sequential feature selection then, finally the obtained statistics was fed to a support vector
machine (SVM) binary classifier to diagnose skin biopsies from patients as either
malignant melanoma or benign nevi. Experiment 5: In this experiment, 42 images were
sampled from microscopic slides of skin biopsy and have been used in the current work.
The algorithm used in these experiments was a 5-fold cross validation test. The experiment
used LIBSVM [310] library for support vector machines, with linear, (RBF) Radial Basis
Function kernel [310], and Polynomial kernel, under each magnification. The melanoma
images that are confirmed by the pathologist are treated as negative images, while the
nevus ones are treated as positive images. Experiment 6: This experiment proposed an
automated system uses a wavelet packet transform (WPT) to extract more features. This
thesis introduces a novel melanoma detection strategy using a hybrid particle swarm based
support vector machine (SSVM) algorithm. The extracted features are reduced by using a
particle swarm optimization (PSO were used to optimize the SVM parameters and finally,
the obtained statistics were fed to a support vector machine (SVM) binary classifier to
diagnose skin biopsies from patients as either malignant melanoma or benign nevi.
CHAPTER 6 Experimental Results and Discussion
152
Table 6.1 Sensitivity and Specificity comparison.
The presented PSO-SVM system resulted in a sensitivity of 94.1 % a specificity of 80.2%
and an accuracy of 87.1 %. The obtained sensitivity and specificity results are comparable
to those obtained by Dermatologists and considerably better than those obtained by less
trained doctors as seen in table 6.1 (quoted from [311]). Consequently, the proposed
system can be considered as a promising method to be used by pathologists for skin cancer
diagnostic. It can serve as a second opinion for the doctor in the skin cancer diagnosis
process.
This experiment investigates skin cancer segmentation using an active contour model, or
snake model, driven by Gradient Vector Flow (GVF). A snake is a parameterized contour
[312] that translates and deforms on the image plane according to the strength of the
image's edges and the internal properties of the contour such as smoothness. This thesis
implements a gradient vector flow (GVF) snake originally developed and discussed in
[313]. The advantage of a GVF snake is its robustness to the initialization of the snake in
[313], [314]. Although a GVF snake is robust to the initialization of the snake, it is still
sensitive to image noise [315]. In order to make the snake insensitive to noise and to be
able to remove the hairs, an Adaptive Filter (Wiener and Median filters) is proposed. After
the noise and hairs are removed, the GVF snake will be used to segment the skin cancer
region. The GVF snake extends the single direction and allows it to still be able to track the
boundary of the skin cancer even if there are other objects near the skin cancer region.
The block diagram in figure 6.1 shows, stage one; the original input colour cancer image,
Stage two converts the colour image RGB into a greyscale image figure 6.2. In stage three,
wiener and median filters are proposed; Wiener filtering is a general way of finding the
best reconstruction of a noisy signal[227]. It gives the optimal way of narrowing off the
CHAPTER 6 Experimental Results and Discussion
153
Figure 6.1 Block Diagram for skin cancer image segmentation
Figure. 6.2 Skin Cancer Image, (a) Original RGB true colour image
(b) Converted greyscale intensity image
noisy components, so as to give the best reconstruction of the original signal. Wiener2 uses
a pixel wise adaptive Wiener method based on statistics estimated from a local
neighborhood of each pixel. Equation 6.1 showed Matlab form[228]:
J = wiener2 (I, [m n], noise) 6.1
Where filters the image using pixel wise adaptive Wiener filtering, using neighbourhoods
of size to estimate the local image and . The
argument, m and n default values is 3. The additive noise (G Gaussian white noise)
power is assumed to be noise. Wiener filtering is best when the basis most cleanly
separates signal from noise, and that’s clear when Wiener filter is applied to the difference
between an image and a smoothed image.
As will The Modified Adaptive Center Weighted Median (MACWM) filter proposed
with an adjustable central weight obtained by partitioning the observation vector
space[233].
CHAPTER 6 Experimental Results and Discussion
154
Figure 6.3. The basic structure of the MACWM filter
The framework of the MACWM filter is illustrated in Figure 6.3. It is composed of four
parts: a median filter, a set of threshold by Fuzzy C-means (FCM), training the centre
weight each block by Least Mean Square (LMS) algorithm, and a decision as to whether
noises exist or not. At first, according to observation vector space (input image); median of
input image will be calculated. We propose FCM algorithm to partition observation vector
space to M block, and related weights to any blocks will be trained by using the LMS
algorithm. The output value y (k) of the MACWM filter (Figure 4.6) at the processed pixel
x (k) is obtained as follows in equation 6.2:
6.2 That the usual output value from a median filter is denoted as m (k) at location k in a filter
window of size 2n + 1 as follows in equation 6.3:
6.3
Where MED means the median operation.
The MACWM filter achieves its effect through the linear combinations of the weighted
output of the median filter and the related weighted input signal. Here, w (k) denotes the
membership function indicating to what extent an impulsive noise is considered to be
located at the pixel x (k). If w (k) = 1, an impulsive noise is considered to be located at
pixel x (k), and the output value of the filter is equal to the output value of the median
filter. If w (k) = 0, an impulsive noise is considered not to be located at pixel x (k), and the
output value is the same as the input x (k); that is, the pixel x (k) is left unchanged. To
judge whether an impulsive noise exists or not at the input pixel x (k), the membership
function w (k) should take a continuous value from 0 to 1. Therefore, the major concern of
the MACWM filter is how to decide the value of the membership function w (k) at the
pixel x (k). The value of K set at 0.2 [316] in the experiment.
CHAPTER 6 Experimental Results and Discussion
155
In general, the amplitudes of most impulses are more prominent than the fine changes of
signals. Thus, the following two variables can be defined to generate the observation
vector[235] . (Partitioning of observation vector space)[233]. In stage four, the filtered
image is then treated by rough segmentation using a thresholding method developed in
[298]. In stage five, Gradient Vector Flow (GVF) snake is proposed. There are two general
types of active contour models in the literature today: parametric active contours[317] and
geometric active contours[318] . In this paper, we focus on parametric active contours,
which synthesize parametric curves within an image domain and allow them to move
toward desired features, usually edges. Typically, the curves are drawn toward the edges
by potential forces, which are defined to be the negative gradient of a potential function.
Additional forces, such as pressure forces[319], together with the potential forces comprise
the external forces. There are also internal forces designed to hold the curve together
(elasticity forces) and to keep it from bending too much (bending forces).
In stage five, Gradient Vector Flow (GVF) snake is proposed, to obtain the final contour
[320]. Particular advantages of the GVF snake over a traditional snake are its insensitivity
to initialization and ability to move into concave boundary regions. A traditional snake is a
curve , that moves through the spatial domain of an image to
minimize the energy functional, Equation 6.4:
6.4
Where and are weighting parameters that control the snake's tension and rigidity,
respectively. and denote the first and second derivatives of with respect
to . The external energy function is derived from the image so that it takes on its
smaller values at the features of interest, such as boundaries.
To add additional flexibility to the snake model, it is possible to start from the force
balance equation directly, as displayed in equation 6.5:
6.5
Balloon models[319] comprise an important example of this approach. In these models
is the sum of the traditional potential forces and pressure (or normal) forces, which
act in a direction normal to the curve. This increases the capture range of an active contour,
but also requires that the balloon be initialized to either shrink or grow.
CHAPTER 6 Experimental Results and Discussion
156
In our first experiment, we computed the GVF field for the line drawing of Figure 6.5(a)
using Comparing the resulting field, shown in Figure 6.5(b) to the potential force
field of Figure 6.4 (b), reveals several key differences. First, the GVF field has a much
larger capture range. It is clear that in order to get this extent using traditional potential
force fields, one would have to use a large in the Gaussian filter. But this would have the
effect of blurring (or perhaps even obliterating) the edges, which is not happening in the
GVF field. A second observation is that the GVF vectors are pointing somewhat downward
into the top of the U-shape, which should cause an active contour to move farther into this
concave region. Finally, it is clear that the GVF field behaves in an analogous fashion
when viewed from the inside. That is, the vectors are pointing toward the boundary from as
far away as possible and are pointing upward into the concave regions (the fingers of the
U-shape) as viewed from the inside.
Figure 6.5 (c) shows the result of applying a GVF snake with parameters and
to the line drawing shown in Figure 5 (a) (using the external GVF field of Figure
5(b)). In this case, the snake was initialized farther away from the object than the
initialization in Figure 6.4 (c), and yet it converges very well to the boundary of the U-
shape. It should be noted that the blocky appearance of the U-shape results from the fact
that the image is only 64 X 64 pixels. The snake itself moves through the continuum
(using bilinear interpolation to derive external field forces which are not at grid points) to
arrive at a sub-pixel interpolation of the boundary.
Figure 6.4. A snake with traditional potential forces cannot move into the concave boundary region.
CHAPTER 6 Experimental Results and Discussion
157
Figure 6.5 A snake with GVF external forces moves into the concave boundary region.
These experiments proposed two methods to find the border of the skin cancer. The first
method is an independent search, which is used to find the seed border points. The second
method is used to track the border based on the nearest border point [321]
Figure 6.6, displays the output of the skin cancer greyscale image, (a) filtered image after
Wiener2 filter (b) filtered image after median filter.
Figure. 6.6. Skin Cancer greyscale image showing: (a) Wiener2 filter has removed the spots effectively (b) noisy image filtered by the median filter
Today, health care institutions produce a large amount of image data which are used in
diagnosis and treatment of patients. As the number of images produced increases,
utilization, and handling of image data are becoming a difficult task for engineers,
scientists, and medical physicists. Researchers say that there are two main concerns in the
field of image processing and analysis applied to medical applications: 1. improving the
CHAPTER 6 Experimental Results and Discussion
158
quality of the image data 2. Featuring of information from medical image data in an
efficient and accurate manner.
In order to demonstrate image quality, we have to start with a basic image processing
overview as defined in Ref.[237]. In this experiment the process of digital images,
described the brightness variation in the image using its histogram. By changing the
histogram, processes that shift and scale the result (making the image brighter or dimmer,
in different ways). Refer to Figure. 6.7 (a) and (b) to see image histogram stretching the
pixel values onto (0, 255). Also thresholding techniques that turn an image from Gray level
to binary were considered. The statistical operations can reduce noise in the image, which
can benefit the feature extraction techniques to be considered later [322].
(a) Without filtering, (b) with filtering.
Figure.6.7 Skin Cancer Image Histogram,
This experiment was carried out on selected 8 (eight) skin cancer images; the images are
distorted with Gaussian noise with mean = 0 and standard deviation = 0.01. The threshold
value is selected to be 0.5 for all the images. The performance of the proposed operators
for edge detection is evaluated and compared with gradient vector flow (GVF) snake;
figure 6.8 illustrates a sample of these images).
In order to compare the segmentation performance of the proposed algorithm, the root-
mean-square error [323]between the input image and the output image of same size MxN
[rows and columns] and the root-mean-square of signal to noise ratio SNRrms used
between the original input image and the restored output image. The root-mean-square
CHAPTER 6 Experimental Results and Discussion
159
error [32]between the original input image and the restored output image
of size MxN [rows and columns] is defined as in equation 6.6:
6.6
The root-mean-square of signal to noise ratio between the original input image
(x, y) and the restored output image of size are defined as in equation 6.7:
6.7
The results displayed in table 6.2 for the proposed operators, the Sobel and the Prewitt
operator showed the best performances (good contrast, edge map strength, and low noise
content) and Robert and Laplacian have poor performance in terms of all indices].
(a) (b)
Figure. 6.8 Skin Cancer greyscale image showing: (a) Sobel segmented image
(b) Gradient vector flow (GVF) segmented image
This experiment stated that the proposed operator displayed the extraction of a good
quality edge map from the grey scale image, and presented a skin cancer image
segmentation algorithm, which classified the most commonly used algorithms. Then these
algorithms have been applied to 8 images. Table 6.2 showed a comparative way: Prewitt is
more acceptable than the other operator, Laplacian is very much sensitive to noise and
Canny operator shown better detection, good response and improving signal to noise ratio.
CHAPTER 6 Experimental Results and Discussion
160
Table 6.2 Comparison between proposed operators (Sobel, Roberts, Prewitt, Laplacian, Canny and Otsu) and Gradient vector flow (GVF) segmented images.
Operators Root mean square error
Signal to noise ratio
Sobel operator 2.3478 0.2081
Roberts operator 2.8630 0.2100
Prewitt operator 1.9687 0.3808
Laplacian operator 1.6930 0.4147
Canny operator 3.7365 0.3257
Otsu operator 1.4214 1.9052
The subjective evaluation of edge detected images shows that the proposed operators
evaluated both the performance and root mean square of signal to noise ratio between the
input image (operators) and the output image (GVF). The advantage of a GVF snake is its
robustness to the initialization of the snake, although it is still sensitive to image noise.
These experiments propose two segmentation methods to identify the normal skin cancer
oriented texture and animation edges and then finding the accuracy by using the three
layers back-propagation neural network classifier (BNN). The experiment results in table
3, displayed the best result with highest overall accuracy is 89.9%. The best BNN is three
hidden layer with 40, 25 and 10 neurons for each hidden layer. The accuracy is increase
with number of neuron in hidden layer. However, number of hidden layer cannot improve
the result but it could reduce the probability of over-fitting [324]. (see section 6.1.2.1.
Experiment Results)
This experiment used a wavelet decomposition method to find better feature extracting
while using one of the more interesting methods, the three layer back-propagation neural
network (BNN) to maintain the classification method. Wavelet decomposition allows
useful cancer’s feature to be extracted from the images without clinical knowledge. In this
system, 2-D wavelet packet is used and the enhanced image in gray scaled as an input.
Since there is lack of published articles to compare different wavelet, seven different
CHAPTER 6 Experimental Results and Discussion
161
wavelets are used in this paper followed by same classification test to compare them and
find out the best results. The ‘bior5.5’ wavelet found it provided the highest accuracy that
achieved 88.5% as displayed in table 6.3. On the other hand, the daubechies wavelet (‘db1’
and ‘db10’) has the most stable experimental record during the testing[324].
Table 6.3 BNN Classification Test with Different Wavelet.
The database images used 448 dermoscopy images collected from the Sydney Melanoma
Diagnostic Centre in Royal Prince Alfred Hospital and internet websites. These images
include digital malignant melanoma images and digital benign melanoma images. The
images were separated into four groups to test the accuracy of each group. The database
images are obtained from different sources and the size of the images is non-standard. The
first step in this process is to resize the image, second apply preprocessing to remove the
fine hairs, noise and air bubbles on the skin and facilitating image segmentation by using
wiener and median filters. These basic operations are usually for pre-processing to improve
display quality. These experiments used a powerful way to enhance the image contrast by
stretching the histogram; these are a very popular technique known as histogram
equalization, which alters pixel values to achieve a uniform distribution.
A post-processing method is used to enhance the shape of the image; a smart contrast
enhancement technique used the histogram equalization (HE) algorithm to stretch the
dynamic range of the image histogram and resulting in overall contrast improvement.
Figure 6.9 shows the result of segmentation by threshold with Gray-scale Histogram
Equalization (GHE) (bottom right) and Local Histogram Equalisation (LHE). The LHE
(bottom left) sharpens the lesion but also reduces the surrounding detail.
CHAPTER 6 Experimental Results and Discussion
162
Figure 6.9. Result of segmentation by threshold with GHE (bottom right) and LHE (bottom left), Top: shows histogram Local and Global results
The segmentation methods used the region of interest (ROI) to extract the interesting
features of melanoma within the border since most of the cancer cells are lump structures.
The border structure provides vital information for accurate diagnosis. Many clinical
features including asymmetry and border irregularity are calculated from the border.
Threshold segmentation is an advanced method for segmentation to compute the intensity
value from grey images. Figure.6.10 shows the result of threshold segmentation where the
white color represents the region area of interest (ROI).
Figure 6.10.Segmentation by Thresholding (ROI)
This experiment used the segmentation to remove the healthy skin from the image and find
the region of interest, usually the cancer cell that remains in the image. A Statistical region
merging (SRM) algorithm is based on region growing and merging. The performance of
SRM has a better segmentation result compared to other popular segmentation methods
[221]. The theory of region merging starts at a seed point and compares its four neighbor
points or pixels. The region is grown when the neighbors pixel share the same homogenous
properties. SRM usually works with a statistical test to decide the merging of regions.
Figure. 6.11 show the result of SRM segmentation where the white color represents the
CHAPTER 6 Experimental Results and Discussion
163
area of interest, which is more accurate in segmentation. It includes more useful areas and
does not put the outer region as part of the area of interest.
Figure 6.11. Segmentation by SRM
Compare two segmentation method, SRM has a better and more accuracy in segmentation.
It includes more useful area and does not put the outer region as part of the area of interest.
However, the accuracy for of neural network as classifier could not support that, since by
using Backpropagation neural network (BNN), it found that the threshold has 0.5% higher
accuracy than SRM in the classification test as illustrate in table 6.4.
Table 6.4. BNN Classification Test for SRM & Thresholding.
Feature extraction is the procedure that takes out the hidden properties or the raw data from
the image that is further used in the classifier. These experiments used a modern method
of feature extraction, wavelet, and curvelet transformation .These approaches, such as
discrete wavelet and discrete curvelet, transform the frequency information at different
scales and locations are obtains [325].
Wavelet decomposition, as explained in sections 5.4.6, can be a repeated process occurring
over several levels. The original signals are decomposed into approximations feeding the
next level of decomposition, creating a decomposition tree. The approach of this feature
extraction is similar to the work in [324]; the image energy is distributed according to the
resolution. Each of the nodes represents the represent one feature, which then can be used
as an input for classification stage. The experiments in the above mentioned work stated
that the second level decomposition can generate 16 nodes or features. Two-dimensional
wavelet packet returns the coefficients in a 2 dimensions matrix. The features are
CHAPTER 6 Experimental Results and Discussion
164
calculated by their mean, maximum, minimum, median, standard deviation and variance.
Therefore, 144 features for each image (16 nodes x 9 features) for 448 images are
produced to be the input for the classification system. This would increase the computation
time without any improvement to the performance [324]. Therefore principal component
analysis (PCA) is used before the training process, the PCA factor chosen = 0.002, to
reduce the 144 features to 60 features. Figure 6.12. Shows the original image and its
transform (wavelet coefficients). This method allows useful cancer features to be extracted
from the images without clinical knowledge. In these experiments, 2-D wavelet packet is
used to enhance the image in grey scaled as an input.
Figure 6.12. Display of the image and its transform (wavelet coefficients)
As explained in [304], Curvelet are waveforms which are highly anisotropic at fine scales,
with effective support obeying the parabolic principle (length width2). Just as for wavelets,
there is both a continuous and a discrete curvelet transform as shown in figure.6.13
Figure. 6.13: A single curvelet with width and length .
CHAPTER 6 Experimental Results and Discussion
165
curvelet have useful geometric features that set them apart from wavelets and combine the
frequency responses of the curvelet at different scales and orientations, a rectangular
frequency tiling that covers the whole image in the spectral domain as shown in Figure
6.14. Thus, the curvelet spectra completely cover the frequency plane and there is no loss
of spectral information as there is with the Gabor filters.
Figure 6.14. Rectangular frequency (basic digital) tiling of an image with 5 level curvelet.
To achieve higher level of efficiency, the curvelet transform is usually implemented in the
frequency domain. The product is then inverse Fourier transformed to obtain the curvelet
coefficients. The process can be described as:
Curvelet transform = IFFT [FFT (Curvelet) x FFT (Image)], and the product from the
multiplication is a wedge.
These experiments calculate the mean and standard deviation for the first half of the total
cells at each scale. The test on melanoma image using the curvelet method showing the
original, noisy and the output of curvelets is displayed in Figure 6.15.
CHAPTER 6 Experimental Results and Discussion
166
Figure 6.15. Curvelet demising image that contains oriented texture and cartoon edges (a) Original (b) Noisy, and (c) Curvelet.
Classification system used to identify if the image is the image if it cancer or not. The
classification system is usually supported by an intelligent classifier, such as a neural
network or support vector machine.
This experiment used 60% of the total 448 images for training, 20% for testing and 20%
for validation. The best result with highest overall accuracy is 70.44% for curvelet with
Back-Propagation Neural Network (BNN) with three layers [40 25 10 neurons] as shown
in table 6.6. The accuracy for neural network as a classifier is increased with the number of
neurons in the hidden layer. However, the number of hidden layers could reduce the
probability of over-fitting; it was found that the curvelet method has 14.5% higher
accuracy than wavelet in the classification test as illustrated in table 6.5.
Table 6.6. Displaying the comparison between wavelet and curvelet features results. Method No of Neurons No of Layers Training Validation Testing Wavelet 40 25 10 3 76.29 55.67 55.94 Curvelet 40 25 10 3 79.76 68.06 70.44
This experiment proposes an automated non-invasive system for skin cancer (melanoma)
detection based on Support Vector Machine classification. The proposed system uses a
number of features extracted from the Wavelet or the Curvelet decomposition of the
grayscale skin lesion images from the original color images. The dataset used include both
digital images and Dermoscopy images for skin lesions that are either benign or malignant.
The recognition accuracy obtained by the Support Vector Machine classifier used in this
experiment is 87.7% for the Wavelet based features and 83.6 % for the Curvelet based
CHAPTER 6 Experimental Results and Discussion
167
ones. The proposed system also resulted in a sensitivity of 86.4 % for the case of Wavelet
and 76.9% for the case of Curvelet. It also resulted in a specificity of 88.1% for the case of
Wavelet and 85.4% for the case of Curvelet. The obtained sensitivity and specificity results
are comparable to those obtained by dermatologists.
The aim was to evaluate the ability of the proposed system to classify digital images of
skin lesions as benign or malignant. The dataset used in the evaluation was a collection of
images obtained from the Sydney Melanoma Diagnostic Centre in Royal Prince Alfred
Hospital and internet websites. In total, the dataset included 448 images divided into four
groups as follows: Group A contains 93 digital images for malignant skin lesions obtained
from the internet; group B contains 142 dermoscopy images for malignant skin lesions
obtained from the Sydney Melanoma Diagnostic Centre; group C and D contain 121 and
92 digital images for benign skin lesions obtained from the internet.
Four experiments were conducted; the data groups for each experiment are varied for the
training and testing or the feature extraction method. The classification technique used for
all of the experiments was the Support Vector Machines (SVM) with Radial Basis
Function Kernel using LIBSVM [326]. I trialed different kernels during my experiments,
in each case, 10-fold cross validation was used to pick the best RBF kernel parameters,
namely, C and . because it demonstrated the best results. For each experiment, 0.8 of the
data was used for training and 0.2 was used for the purpose of testing. The experiments
conducted are described below. The data used for the computation of the specificity and
sensitivity is based on dataset. In the experiment, used 60% of the total images for training,
20% for testing and 20% for validation.
Test 1 included dataset of 448 images both digital skin images and dermoscopy images
and was used for the evaluation purposes and Wavelets ‘db4’were used for the feature
extraction step. Three evaluations were conducted based on the number of features
extracted from the second level wavelet decomposition which resulted in 16 two-
dimensional matrices. The first used 9 features per matrix (calculated as a combination of
the mean, maximum, minimum, median, standard deviation and variance of each two
dimensional matrix). The second used the same 9 features together with L1 and L2 norms.
The third used the previously mentioned 11 features and the variances of the R, the G and
the B layers of the image to include the color effect. The results of test 1 are given in table
CHAPTER 6 Experimental Results and Discussion
168
6.7. It can be seen from this table that the cross validation accuracy increases with the
number of features and that including the colour features improved the cross validation
accuracy by 6.7%.
Test 2, the dermoscopy data was excluded. Only digital skin images (305 images) were
used for the evaluation purposes. Similar to test 1, Wavelets ‘db4’were used for the feature
extraction step. The wavelet function Daubechies ‘db4’ was used because it has the most
stable experimental record during the testing.Three evaluations were conducted based on
the number of features extracted from the second level wavelet decomposition which
resulted in 16 two-dimensional matrices. The first used 9 features per matrix (calculated as
a combination of the mean, maximum, minimum, median, standard deviation and variance
of each two dimensional matrix). The second used the same 9 features together with L1
and L2 norms. The third used the previously mentioned 11 features and the variances of
the R, the G and the B layers of the image to include the colour effect. The results of test 2
are shown in table 6.4. It can be seen from table 6.7 that the cross validation accuracy
increased with the number of features. It can be also noticed that the specificity, sensitivity
and the testing accuracy increased dramatically with the inclusion of the colour features.
The testing accuracy resulting when including the colour features is 87.7 % with around
11% increase compared with the same case in table 6.8 with the dermoscopy images
included in the dataset. The trend of improved testing accuracy with color features
observed in test 2 was not noticed in test 1 as the image set used included two types on
images with different characteristics namely the digital skin images and the dermoscopy
images.
Test 3 used the 448 images which include both digital skin images and dermoscopy images
for the evaluation. Curvelet were used for the feature extraction step. Two evaluations
were conducted based on the number of features obtained from the Curvelet decomposition
of the image. The first used 52 features (calculated using the mean and standard deviation).
The second used the same 52 features together with the variances of the R, the G and the B
layers of the image to include the colour effect. The results of test 3 are given in table 6.9.
This table showed that the cross validation accuracy increases with the number of features
and that including the colour features improved the cross validation accuracy by 4.4%.
Test 4, the dermoscopy data was excluded. Only digital images (305 images) were used for
the evaluation purposes. Similar to test 3, curve lets were used for the feature extraction
CHAPTER 6 Experimental Results and Discussion
169
step. Two evaluations were conducted based on the number of features used. The first used
52 features (calculated from the mean and standard deviation of the Curvelet coefficients).
The second used the same 52 features together with the variances of the R, the G and the B
layers of the image to include the colour effect. The results of test 4 are given in table 6.10.
This table showed that the cross validation accuracy increased with the number of features.
It can be also being noticed that the specificity, sensitivity and the testing accuracy
increased with the inclusion of the colour features. The resulting testing accuracy with the
inclusion of with the inclusion of the colour features is 83.6 % with around 17.6 % increase
compared with the same case in table 3 with the dermoscopy images included in the
dataset. The dramatic improvement in the specificity, sensitivity and the testing accuracy
for test 4 compared to test 3 can be justified by understanding that the dataset used for test
4 was more homogenous as it included only digital images for the skin lesions as opposed
to test3 where the dataset included two types on images with different characteristics
namely, the normal digital images and the dermoscopy images. The inclusion of two types
on images with different characteristics reduced the ability of the SVM to generalize as is
clear from the fact that the cross validation accuracy results were considerably higher than
the testing accuracy for test 3.
In general, it can be said that the inclusion of the colour features improved the testing
accuracy and that the Wavelet extracted features gave better performance (87.7 %)
compared with those found by Curvelet extracted features (83.6 %).
Table 6.7. svm classification for all the dataset (including dermoscopy images) with features extracted from wavelets
CHAPTER 6 Experimental Results and Discussion
170
Table 6.8. svm classification excluding dermoscopy images with features extracted from wavelets
Specificity
Sensitivity Testing
Accuracy
10-fold Cross Validation Accuracy
RBF Kernel Parameters
Wavelet 9 Features
76.6
64.7 74.1 78.6 C=362.0
Wavelet 11 Features
81.0
65.2 76.5 84.8 C=4.757
Wavelet and Color Features
88.1
86.4 87.7 85.3 C=1024
Table 6.9. svm classification for all the dataset (including dermoscopy images) with features extracted from curvelets
Specificity
Sensitivity Testing
Accuracy 10-fold Cross Validation Accuracy
RBF Kernel Parameters
Curvelet Features
68.6
63.0
65.2
76.9
C=13.45
Curvelet
and Color
Features
76.9
61.9 66.3 81.3 C=11.31
Table 6.10. svm classification excluding dermoscopy images with features extracted from curvelets
Specificit
y
Sensitivity Testing Accuracy
10-fold Cross Validation Accuracy
RBF Kernel Parameters
Curvelet Features
83.3
69.2
80.3
76.6
C=2
Curvelet and Color Features
85.4
76.9
83.6
78.6
C=861.1
The presented system resulted in a sensitivity of 86.4 % for the case of wavelet and 76.9%
for the case of curvelet. It also resulted in a specificity of 88.1% for the case of Wavelet
and 85.4% for the case of curvelet. The obtained sensitivity and specificity result are
comparable to those obtained by dermatologists as seen in table 6.11 from [310].
Consequently, the proposed system can be considered as a promising non-invasive method
to be used for skin cancer diagnosis that does not require well trained practitioners as for
the case of Dermoscopy images. It can serve as a second opinion for the doctor in the skin
cancer diagnosis process.
CHAPTER 6 Experimental Results and Discussion
171
Table 6.11. sensitivity and specificity statistics [310]
Sensitivity Specificity
Experts 90 % 59 %
Dermatologists 81 % 60 %
Trainees 85 % 36 %
General Practitioners 62 % 63 %
The next part based on pathological images is; melanoma is a pathology clinical lesion is
irregularly shaped, an asymmetrical lesion with varying colour with a history of recent
change in size, shape, colour or sensation. Melanoma may arise within an existing benign
lesion. Most melanomas have an initial radial growth phase within the epidermis and
sometimes within the papillary dermis, which may be followed by a vertical growth phase
with deeper extension. The experiments 1, 2 and 3 concludes that there are some possible
factors to improve the accuracy of detecting malignant melanoma by having a higher
number of images for training of the SVM network.
The complexity and the parameter adjustment can be a serious problem in several
industrial computer vision applications; it had been proposed a spatially adaptive median
filter (SAMF) would be more efficient for image processing [327]. This filter (SAMF)
applied widely as an advanced method compared with standard median filtering is used
[328]. Color-based segmentation is one of the most widely investigated research areas in
pathological image analysis. Fundamentally, there are several objects and each object has
several features and this algorithm has been applied to classify the objects based on the
features [110].
One of the methods had been used in this thesis to find the features is by generalizing the
k-mean clustering into n features, and to define the centroid as a vector where each
component is the average value of that component. Each component represents one
features [329] .
The proposed system used a number of features extracted from histo-pathological images
of skin lesions through image processing techniques which consisted of a spatially adaptive
color median filter for filtering and a k-means clustering for segmentation. The extracted
features were reduced by using sequential feature selection and then classified by using
CHAPTER 6 Experimental Results and Discussion
172
support vector machine (SVM) to diagnose skin biopsies from patients as either malignant
melanoma or benign nevi.
The results on figure 6.16 show, original image, image after adaptive median filter, gray
scale image, histogram out, adjust image intensity, and output image using histogram
equalization.
This experiment is a new application based on histo-pathological images of skin lesions
that required finding out new features and the correlation of the reduced numbers of
features and getting better accuracy. It was required to modify many of the mentioned
techniques to make them work for such application. The higher accuracy of diagnosis
proposed work is calculated and displayed.
Figure. 6.16. The results showed, original image, image after median filter, gray scale
image, histogram out; adjust image intensity, and output image using histogram
equalization
In this experiment, 42 images were sampled from microscopic slides of skin biopsy and
have been used in the current work.
The algorithm used in these experiments was a 5-fold cross validation test. The experiment
used LIBSVM [310] library for support vector machines, with linear, (RBF) Radial Basis
Function kernel [310], and Polynomial kernel, under each magnification. The melanoma
CHAPTER 6 Experimental Results and Discussion
173
images that are confirmed by the pathologist are treated as negative images, while the
nevus ones are treated as positive images. Researchers, in [330] classifying breast tissue
found 67.1% accuracy in labeling nuclei as benign or malignant; the classification of
histopathology images presented in [310] resulted in 88.7% accuracy in the diagnosis of
lung cancer, 94.9% accuracy in the typing of malignant mesothelioma, and 80.0% to
82.9% accuracy in the diagnosis of malignant mesothelioma for Fulgent stained lung
sections. Keenan et al. [310] reported accuracies of 62.3% - 76.5% in the grading of H&E
stained cervical tissue. Skin cancer is a fast developing disease of modem society, reaching
20% [310] increasing of diagnosed cases every year. Dermatoscopic methods used for
diagnostics are non-invasive and require a great deal of experience to make correct
diagnosis. As reported in [310], only experts arrive at 90% sensitivity and 59% specificity
in skin lesion diagnosis; while for less trained doctors, these figures show a significant
drop to till around 62%- 63% for general practioners, as mentioned on Table 6.3.
In this experiment, the images are sampled and split into a training set and a test set
using MATLAB software. The testing process is done for 42 images (28 benign and & 14
melanoma). These 42 images were are fed to the proposed SVM network In the
experiment, used 60% of the total images for training, 20% for testing and 20% for
validation. The performance of the algorithm was evaluated by computing the percentages
of Sensitivity (SE), Specificity (SP) and Accuracy (AC) [331]. From our experimental
results CPU time faster than experiment 3 [22], from our experimental results CPU time =
9 min., which is 17 times faster than the old experiment[19] ,( CPU time = 156 min) and
produce more accurate classification results, than physicians do.
Table 6.13 shows the average resulting SE, SP and AC of 5 fold validation of the
proposed networks by using LIBSVM. This experiment used RBF kernel with
gamma=0.l25 and C=1. The results showed this to be one of the best results comparing
with expertise / physicians results displayed on table 6.12. The obtained accuracy of the
system is 88.9 % whereas the sensitivity and specificity were found to be equal to 87.5 %
and 100% respectively. By comparing the results on tables 6.12 and 6.13, it is concluded
that the proposed system gives faster and more accurate classification of skin tumors than
physicians do. This experiment concludes that there were some possible factors to improve
the accuracy of detecting malignant melanoma by having a higher number of images for
training of the SVM network.
CHAPTER 6 Experimental Results and Discussion
174
Table 6.12. Described sensitivity and specificity
Sensitivity Specificity
Experts 90% 59% Dermatologists 81% 60%
Trainees 85% 36% General practitioners 62% 63%
Table 6.13. show result after training or the (svm) network. using (sfs) technique
No of images Sensitivity Specificity Accuracy 42 87.5 100 88.9
Early detection of malignant melanoma remains the key factor in reducing mortality from
skin cancer. Recognizing the importance of this issue in the 1990s, dermoscopy enabled
the recognition of new subsurface features to differentiate between malignant and benign
pigmented lesions. During the last decade, new computer-based technologies have
improved diagnostic sensitivity and specificity and may result in optimizing lesion
selection for biopsy and pathology review. Even with all of the advances in melanoma
diagnosis, timely recognition, detection, and rapid treatment of melanoma remain critical.
Although pathologic examination remains the gold standard for diagnosis, this cancer has
the potential to be diagnosed through non-invasive approaches because of its affecting the
skin location. From the development of the ABCDs through current efforts that use
complex computer algorithms and genetic markers, a clinician’s ability to detect melanoma
in its earliest form has been improved. However, a “good clinical eye” is still fundamental
to selecting the lesions for evaluation among the sea of those that are widespread. As
current approaches are refined and new techniques are developed, the improved ability to
diagnose this cancer will hopefully enhance reaching the goal of reducing melanoma
mortality[191].This experiment used a novel methodology for automatic feature extraction
from histo-pathological images and subsequent classification. These techniques added
Gabor filter bank to winner and adaptive median filters to improve diagnostic accuracy.
Then Histogram equalization was used to enhance the contrast of the images. The extracted
features were reduced by using sequential feature selection then; finally the obtained
CHAPTER 6 Experimental Results and Discussion
175
statistics was fed to a support vector machine (SVM) binary classifier to diagnose skin
biopsies from patients as either malignant melanoma or benign nevi.
This experiment used forty two (42) images sampled from microscopic slides of skin
biopsy. MATLAB R2013b software was used; the features were carried out to generate
training and testing of the proposed SVM. , this 42 images (28 benign images and 14
melanoma images) were sampled and split into a training set (mixed 28 images) and test
set mixed 14 images to be used as input to the proposed SVM network. The performance
of the algorithm was evaluated by computing the percentages of Sensitivity (SE),
Specificity (SP) and Accuracy (AC). Table 2 shows the average resulting SE, SP and AC
of 2 fold cross validation of the proposed networks by using L1BSVM. In this study used
RBF kernel with gamma=0.125 and C=10. We did many trials to improve our
experimental results; these experiments present a new application based on histo-
pathological images of skin lesions that required finding out new features and the
correlation of the reduced numbers of features and getting better accuracy. Measuring the
many of techniques to make them work for such an application. The results which are
displayed on table 6.14 in which the SFS technique was used technique produces one of
the good and acceptable results compared with dermoscopy and pathologic diagnosis
results table 6.16 and also with the results displayed on table 6.15, which shows the results
without using (SFS) technique.
Table 6.14. the result after training of the (svm) network with using (sfs) technique.
No of images Sensitivity Specificity Accuracy 42 76 100 81
Table 6.15. the result after training of the (svm) network without using (sfs) technique.
No of images Sensitivity Specificity Accuracy 42 55.6 80.6 68.1
Table 6.16 Comparison of different algorithms for dermoscopic and pathologic
diagnosis
SENSITIVITY SPECIFICITY DIAGNOSTIC ACCURACY
Pattern analysis 85% 79% 71% ABCD 84% 75% 76%
7-Point checklist 78% 65% 58% CASH 98% 68% —
Menzies 85% 85% 81% SolarScan 91% 68% —
SVM using sfs technique 87.5% 100% 88.9%
CHAPTER 6 Experimental Results and Discussion
176
Several different algorithms for dermoscopy diagnosis are used to differentiate melanoma
from benign melanocytic lesions: pattern analysis; ABCD rule of dermoscopy; 7-point
checklist; colour, architecture, symmetry, and homogeneity (CASH); and Menzies method
(Table 6.16)[332] [336] [333]. Based on the results of the Consensus Net Meeting on
Dermoscopy, sensitivity varied between 82.6% and 85.7%, and specificity varied between
70% and 83.4% for distinguishing melanomas from benign lesions[337]. Although pattern
analysis had the highest sensitivity, specificity, and diagnostic accuracy for melanoma
detection when performed by dermatologists, the Menzies method proved better when used
by nondermatology physicians[334]. Clinicians may be discouraged from using
dermoscopy because they feel it is too time consuming.
SolarScan is an automated instrument (Polartechnics, Sydney, Australia), SolarScan was
found to give a sensitivity of 91% and specificity of 68% for melanoma, which was
comparable to that of expert clinicians[338]
The pathologic results obtained in this experiment are comparable to those obtained by
different algorithms for dermoscopic as seen in table 6.16 consequently; the proposed
system can be considered as a promising method to be improved by increasing the number
of images to be used later by pathologists for skin cancer diagnostic.
This experiment proposed an automated system uses a wavelet packet transform (WPT) to
extract more features. This thesis introduces a novel melanoma detection strategy using a
hybrid particle swarm based support vector machine (SSVM) algorithm. The extracted
features are reduced by using a particle swarm optimization (PSO were used to optimize
the SVM parameters and finally, the obtained statistics were fed to a support vector
machine (SVM) binary classifier to diagnose skin biopsies from patients as either
malignant melanoma or benign nevi.
The support vector machine (SVM) classifier is used widely in bioinformatics, due to its
high accuracy, ability to deal with high-dimensional data and in this syntax diverse sources
of data [310] [339] [340]. It is possible that the data found from the RBF mapping is more
likely to be correctly classified by SVM than from the other kernel function mappings.
Furthermore, SVM-linear could be a special case of SVM-RBF [339] [341], which means
CHAPTER 6 Experimental Results and Discussion
177
that SVM-RBF has more possibilities to obtain a better performance than SVM-linear. The
better performance of the SVM based system which employs RBF is confirmed in another
application [310] [342].
This thesis used in total 283 features (six features extracted from histogram, twenty one
features extracted from co-occurrence matrix for each channel and 255 features
extracted from Wavelet Packet Transform (WPT) see chapter 5). WPT was used to
implement the feature extraction process. Variability appears to be what most separates
malignant melanoma from benign nevi, therefore the best approach at feature extraction
would be to retain as much of the data variability as possible [308]. Wavelet analysis looks
at these changes over different scales which should detect whole lesion changes such as
texture, color, and local changes like granularity. The WPT is a generalized version of the
wavelet transform: the high-frequency part also splits into a low and a high frequency part;
this produces a decomposition tree as shown in the Figure. 6.17. The WPT provides a high
dimensional feature vector thus providing more information about the images. However,
the WPT complicates the analysis process as the high dimensionality of the feature vector
causes an increase in the learning parameters of the pattern classifier, and the convergence
of the learning error deteriorates. Consequently, dimensionality reduction played an
important role before applying the feature vector to the pattern classifier. The extracted
features were reduced by using a particle swarm optimization (PSO) [343]. This was used
to optimize the SVM parameters as a feature selection and finally, the obtained statistics
were fed to a support vector machine (SVM) binary classifier to diagnose skin biopsies
from patients as either malignant melanoma or benign nevi. The obtained classification
accuracies show better performance in comparison to similar approaches for feature
extraction.
CHAPTER 6 Experimental Results and Discussion
178
Figure. 6.17. - A Wavelet Packet decomposition tree
The proposed image processing approach for computer-aided melanoma detection has been
presented. Thesis employs the algorithm and the input of pathology images. has
been developed with an output of +1, which means that a melanomas event is happening,
and -1, which means that there is no melanoma (benign).
One of the main parameters of SVM is C that constrains the vector margin as explained
previously in section 5.6. The incensement of does not always provide improvement of
the performances of the algorithms in the melanoma detection. The reason could be
that theoretically an increasing value allows more points to cross the supporting
hyperplanes of . This could reduce the generalization performance of the detection
algorithm. The increase in value from to decreases the testing sensitivities in
terms of geometric means. On the other hand, this increase in raises the training
performances in all terms (sensitivities, specificities and geometric means). The increase of
the training performances (while the testing performance decreases) means that the
generalization of the algorithm becomes worse.
CHAPTER 6 Experimental Results and Discussion
179
The other thing is the modification in the weight factor. The modification of or is
similar to modification of in the associated class. Modification of means
modification of for the data points in the nonmelanomic (benign) class, and modification
of means modification of for the data points in the melanomic class. The optimal
parameters have been found in the optimization stages. The optimal is higher
than . In the optimization, the weight factor is automatically tuned to prevent a low
sensitivity. In other words, the automatic selection of the weight factor performs well.
Our aim is to evaluate the ability of our proposed system to classify Histo-pathological
images of skin lesions as benign or malignant. The dataset used in our evaluation is
obtained from the Southern Pathology Laboratory in Wollongong NSW, Australia. It
includes 79 Histo-pathological images (29 benign images and 50 melanoma images).
Two experiments were conducted. They differ from each other in the feature selection
method employed and the way the optimal SVM parameters are picked (C and ). The
classification technique used for all of the experiments is SVM with Radial Basis Function
Kernel using LIBSVM [344]The algorithms were implemented using MATLAB R2013b.
For each experiment, 0.8 of the data was used for training and 0.2 was used for the purpose
of testing.
The melanoma images that are confirmed by the pathologist are treated as negative images,
while the nevus ones are treated as positive images. The performance of the algorithms is
evaluated by computing the percentage Sensitivity (SE), Specificity (SP) and Accuracy
(AC) using equations 6.8, 6.9, 6.10, respectively:
X 100 (6.8)
X100 (6.9)
(6.10)
Where TP is the number of true positives, TN is the number of true negatives, FN is the
number of false negatives, and FP is the number of false positives.
CHAPTER 6 Experimental Results and Discussion
180
In the first experiment, SFS was used to choose the most relevant features that result in the
best performance of the SVM. 5-fold cross validation was used to pick the best RBF
kernel parameters (C=10 and =0.125). The results of experiment 1 are shown in table
6.17.
PSO-SVM arrangement was used in the second experiment to choose the most relevant
features and the optimal RBF kernel parameters that optimize the performance of the
SVM. 5-fold cross validation was used within the PSO-SVM arrangement. The resulting
RBF kernel parameters are C=3525.0051 and =0.0084732). The results of experiment 2
are shown in table 6.18.
Comparing the results in table 6.17 with those in table 6.18, it can be seen the
performance of the PSO-SVM arrangement for feature selection and SVM parameter
optimization is considerably better than that obtained when using SFS for feature selection.
An improvement of 10 % can be noticed in the values of the Sensitivity, the Specificity
and the Accuracy. This highlights the success of the PSO-SVM arrangement.
The presented PSO-SVM system resulted in a sensitivity of 94.1 % a specificity of 80.2%
and an accuracy of 87.1 %. The obtained sensitivity and specificity results are
comparable to those obtained by Dermatologists and considerably better than those
obtained by less trained doctors as seen in table 6.19 (quoted from [345] )Consequently,
the proposed system can be considered as a promising method to be used by pathologists
for skin cancer diagnostic. It can serve as a second opinion for the doctor in the skin cancer
diagnosis process.
Table 6.17 The Result after Training of the (SVM) Network, with using (SFS)Technique.
No of images Sensitivity Specificity Accuracy 79 83.6 70.7 77.4
Table 6.18 The Result after Training of the (SVM) Network. with using (SVM+WLG+PSO)
No of images Sensitivity Specificity Accuracy
CHAPTER 6 Experimental Results and Discussion
181
Table 6.19 Described Sensitivity and Specificity [31]
Sensitivity Specificity Experts 90 % 59 % Dermatologists 81 % 60 % Trainees 85 % 36 % General practitioners 62 % 63 %
Figure. 6.18 shows the melanoma detection employing SVM (SVMR, SVMP, and SVML); the input is pathology images and the output is melanoma (-1) or benign (+1).
6.3
where:
• TP (true positive) is the number of inputs which correspond to melanoma classified
as melanoma.
• FP (false positive) is the number of inputs which correspond to benign classified as
melanoma
• TN (true negative) is the number of inputs which correspond to benign classified as
benign.
• FN (false negative) is the number of inputs that correspond to melanoma classified
as benign.
The geometric mean (gm) is suitable for indicating the performance of a case with
imbalanced data [346]. The imbalanced data means that the data point number of a class is
significantly more than that of another class, as is the data in this research.
CHAPTER 6 Experimental Results and Discussion
182
The performance of the detection algorithms was measured in terms of sensitivity,
specificity and geometric mean, instead of a root mean square error or [347]. Using
sensitivity the correct detection of melanomic class data can be known, and using
specificity the correct detection of nonmelanomic (benign) class data can be seen. On the
other hand, only provides one value for the correct detection of both classes, instead
of one value for each class.
The final results in experiment 6 displayed melanoma detections using the swarm based
algorithms. with different kernel functions have been compared.
Furthermore, the detection using has also been compared with the swarm based
multiple regression and the algorithms without the swarm optimization.
The performance of kernel function) is better compared to the
other three , which employ polynomial, sigmoid and linear kernel functions. It is
possible that the data found from the mapping is more likely to be correctly classified
by than from the other kernel functions mappings. Furthermore, could
be a special case of [296], which means that has more
possibilities to obtain a better performance than . The better performance of
the based system which employs is confirmed in another application [263].
The proposed swarm optimization could be suitable to prevent a low sensitivity in
melanoma detection. A high sensitivity of a disease detector means it has a high true
positive. The high true positive in melanoma detection means that the melanomas events
can be detected properly. (more detailed information in Appendix H: Abbreviation)
In the fitness function for the optimization section 5.8.3 and (Eq. 5.53), was set to 0.58.
This value was determined through several experiments. The important thing of value
is that the coefficient for the sensitivity is set higher than the coefficient for the
specificity. By this setting a low sensitivity could be avoided. This setting is necessary
because the data number of the melanomas class is far lower than the nonmelanomic
(benign) class.
This difference has the risk of low sensitivity. However, if is too high, the sensitivity
could be very high and the specificity could be very low. Therefore, the value should be
more than 0.50, but not too high[348]. (refer to section 5.8.3, fitness function for the
optimization).
CHAPTER 6 Experimental Results and Discussion
183
The computational time required by the optimization using depends on the iteration
number provided and the time consumed in every iteration. Every iteration creates
training to make models; is the number of particles of . In this thesis, the
iteration number is 200 and the particle number is 50.
Hence, after 200 iterations the optimization terminates. Another termination principle is
the rises of the global best during ; if the rises are less than a defined value,
the optimization terminates. It means that if the optimal values do not change significantly
during m iteration, the optimization terminates. (More detailed information in applications
E).
The proposed swarm based outperforms the swarm based multiple regression ( )
as showen in experiment 6. The optimization methods of these two algorithms are the
same, in which their parameters are optimized using the . One difference of the two
algorithms in finding the decision function is that considers maximizing the margin
between the two nearest points of the two classes in , whereas does not. The
maximization of the margin could result in better generalization.
This chapter ‘Experimental Results and Discussion’ is organized and structured in
describing the experiments based on digital and pathological images; experiment 1
investigates skin cancer segmentation using an active contour model, or snake model,
driven by Gradient Vector Flow (GVF); experiment 2 proposed two segmentation methods
to identify the normal skin cancer oriented texture and animation edges followed by
finding the accuracy by using the three layers back-propagation neural network classifier
(BNN); experiment 3 proposed an automated non-invasive system for skin cancer
(melanoma) detection based on support vector machine classification. The proposed
system uses a number of features extracted from the wavelet and the curvelet of
decomposed greyscale skin lesion images from the original color images; experiment 4
proposed a system used a number of features extracted from histo-pathological images of
skin lesions through image processing techniques which consisted of a spatially adaptive
color median filter for filtering and a k-means clustering for segmentation. The extracted
CHAPTER 6 Experimental Results and Discussion
184
features were reduced by using sequential feature selection and then classified by using
support vector machine (SVM) to diagnose skin biopsies from patients as either malignant
melanoma or benign nevi; experiment 5 proposed a novel methodology for automatic
feature extraction from histo-pathological images and subsequent classification. These
techniques added Gabor filter bank to winner and adaptive median filters to improve
diagnostic accuracy. Then Histogram equalization was used to enhance the contrast of the
images. The extracted features were reduced by using sequential feature selection then.
Finally the obtained statistics was fed to a support vector machine (SVM) binary classifier
to diagnose skin biopsies from patients as either malignant melanoma or benign nevi.
Experiment 6 proposed an automated system using a wavelet packet transform (WPT) to
extract more features. This thesis introduces a novel melanoma detection strategy using a
hybrid particle swarm based support vector machine (SSVM) algorithm. The extracted
features are reduced by using a Particle Swarm Optimization (PSO) were used to optimize
the SVM parameters and finally, the obtained statistics were fed to a support vector
machine (SVM) binary classifier to diagnose skin biopsies from patients as either
malignant melanoma or benign nevi.
CHAPTER 7 Experimental Results and Discussion
185
CHAPTER 7
Conclusion and Future Work
7.1. Conclusion
Skin cancer is increasing and affects many people in different countries in the world
including Australia. Malignant melanoma is the deadliest type of skin cancer that can be
treated successfully if it is detected early. Early intervention will lead to a better survival
rate. Since the clinical observation of a melanoma is subject to human error, early detection
can be enhanced by utilising an automated process. In this thesis, an automatic skin cancer
detection system had been developed to assist doctors in the accurate detection of skin
cancers while reducing the computational time. The database images were enhanced
through an image processing procedure. This study intended to show that it was possible,
with some common utilization hardware – a simple digital camera – and some well-
structured software, to reach results that would increase the possibilities of accurate
detection of malignant skin lesions and, consequently, increase the possibilities of recovery
from skin cancer, thus, also increasing life expectancy of the people affected by this
disease.
Dermatology centers are very specialized work environments that are not always available
in small clinical institutions. Up to now, all the available systems rely on rather
sophisticated and expensive equipment, not directed or even accessible to the general
public. Regardless of the cost associated with the available systems, most of them are
nothing but simple image storage solutions. Those few that implement classification
methods, reached results that are much less than satisfactory.
These considerations made it obvious for the author that, along the research process,
various difficulties would arise. This became a fact, namely in were image capturing
conditions were concerned. These varied environmental conditions create several problems
to the feature extraction system, particularly in the detection of edges, which is, in fact, the
basis for the classification process.
CHAPTER 7 Experimental Results and Discussion
186
The aim of this thesis is to develop an automated process for the rapid and accurate
detection of a benign (safe) or dangerous (melanoma) within a smaller perimeter of error
than that which is achievable by clinical observation. A high performance computer aided
diagnostic system is developed to assist physicians and avoid misdiagnosis.
The common approaches to skin lesion early detection include different stages of pre-
processing, segmentation, feature extraction and selection, and classification. Since the
output of each stage is the input of the next stage, all stages have an important role to avoid
misdiagnosis.
Image enhancement is one of the most important issues in low-level image processing. Its
purpose is to improve the quality of low contrast images. Therefore, the underlying
principle of the enhancement is to enlarge the intensity difference between objects and
surroundings, so that the image is not apparently distorted.
The histogram which has been used in the thesis provides much evidence to represent the
characteristic of the images. The proposed multi-peak Generalized Histogram Equalization
(GHE) method described in described in chapter 4.3.6 is very effective not only in
enhancing the entire image but also in enhancing the image texture. It makes the change of
the order of grey levels of the original image completely controllable.
Image segmentation is a critical and essential component and is one of the most difficult
tasks in image processing, pattern recognition, and determines the quality of the final
analysis. Segmentation subdivides an image into its principal regions or objects.
Segmentation it an essential issue in digital image processing which is used for image
description and classification.
Hundreds of features might be derived from an image. However not all of the features are
suitable for mass classification. Too many irrelevant features will not only make the
classifier complicated, but will also reduce the accuracy of the classification. The most
important issue is to select features that are able to represent the characteristics of masses
in the skin cancer images, and based on these features, the malignant mass can be
significantly discriminated from the benign masses by the classifier. The lump will be
separated from normal skin and useful information will be extracted from the image. Then,
it will pass through classifiers for training and testing to define whether it is malignant
melanoma or benign.
CHAPTER 7 Experimental Results and Discussion
187
Support Vector Machines are a suitable approach to data modelling. They combine
generalisation control with a technique to address the problem of dimensionality.
The detection of melanoma using a swarm-based SVM (SSVM) algorithm has been
conducted. In SSVM, a particle swarm optimization (PSO) was used to optimize the SVM
parameters. A fitness function was defined in the optimization to find a high performance
of the melanoma detection, especially in the sensitivity.
Different kernel functions were investigated to find the suitable kernel function for the
SSVM which incomes a good performance in the melanoma detection. The SSVM which
used the RBF kernel function performs well with 94.1 %, 80.2 % and 87.1 % in terms of
sensitivity, specificity and geometric means. The thesis outcome results are very
significant they are one of the best results in this type of research as seen in table 6.18
describing sensitivity and specificity in chapter 6, section 6.1.6.1, pg. 179. results and
discussion for experiment 6. By Comparing the results in table 6.17 with those in table
6.18, it can be seen the performance of the PSO-SVM arrangement for feature selection
and SVM parameter optimization is considerably better than that obtained when using SFS
for feature selection. An improvement of 10 % can be noticed in the values of the
Sensitivity, the Specificity and the Accuracy. This highlights the success of the PSO-SVM
arrangement
7.2. Future Research
Future research aims to develop hybrid artificial intelligence techniques employing the
following genetic algorithms; hybrid of fuzzy inference system (FIS), Ant Colony
Optimization (ACO), called PSO-ACO, Ontology, Contourlet Transform and others in
multifunction pattern recognition systems.
Looking forward to conduct further experiments and to test this algorithm on a large
dataset having different pathologies and also with different skin colours.
Future work will propose the implementation of this approach and at the end, with strategic
selection of different fusions tested within this report, we can better give global diagnostics
for any given pathology.
CHAPTER 7 Experimental Results and Discussion
188
This chapter conclusion and future research has presented the contributions to melanoma
detection technology. The models of melanoma detection have been developed and
examined. In addition, the hybrid algorithms with the base of SVM were introduced and
tested for melanoma detection. The main algorithms are support vector machine (SVM),
and a swarm optimization technique (PSO).
The detection of melanoma using a swarm-based SVM (SSVM) algorithm has been
conducted. In the SSVM, all the possible combinations of the SVM parameters are tested
for the input. The SSVM performs better than SVM. In the SSVM, and SVM are combined
and the SVM parameters are optimized using a swarm technique. In terms of
computational time, the two algorithms differ in the off-line Optimization stage. The
computational time can be controlled so that the time could still be acceptable. In the
experiment, the SSVM performs better than the SVM algorithm, with the sensitivity,
specificity and geometric mean of 94.1 %, 80.2 % and 87.1 %, respectively.
we plane to conduct further experiments and to test this algorithm on a large dataset with
selection of different fusions tested within this report, to present better diagnostics for any
given pathology.
Appendices Experimental Results and Discussion
189
Appendices
A. Appendix A: Glossary of Pathology Terminology
• Actin: a protein found especially in microfilaments (as those comprising
myofibrils) and it is active in muscular contraction, cellular movement, and
maintenance of cell shape.
• Adenocarcinoma: a malignant tumour originating in glandular epithelium.
• Fine-needle aspiration cytology (FNAC): a diagnostic procedure used to investigate
superficial (just under the skin) lumps or masses. In this technique, a thin, hollow
needle is inserted into the mass for sampling of cells. These cells are examined
under a microscope after being stained. Figure A1 & A2
Figure A1, showed a physician's hands are seen performing a needle biopsy to determine nature of lump either fluid-filled cyst or solid tumour
[This image was released by the National Cancer Institute, an agency part of the National Institutes of Health, with the ID 1973 (image)]
Figure A2, the diagnostics showed: Micrograph of a needle aspiration biopsy specimen of a salivary gland showing adenoid cystic carcinoma. (Source: Pap stain. MeSH D044963).
Appendices Experimental Results and Discussion
190
• Benign: not causing death / disease without cancer/ not cancerous. Figure A3
Figure A3, showed a Normal Epidermis and Dermis with Intradermal Nevus 10x-cropped. (Source, Kilbad).
• Brightfeld microscopy: This is the simplest of a range of techniques used for
illumination of samples in light microscopes. It is a simple technique and that made
it a quite popular. The typical appearance of a bright-field microscopy image is a
dark sample on a bright background. Figure A4.
Figure A4, this image shows a crossection of the vascular tissue in a plant stem. Showed as an example for bright field micrograph. (Source: Wikipedia).
• Carcinoma: is a type of cancer.
• Cellularity: is the state of a tissue or other mass with regard to the number of
constituent cells.
• Confocal microscopy: is an optical imaging technique used to increase optical
resolution and contrast of a micrograph by using point illumination and a spatial
pinhole to eliminate out-of-focus light in specimens that are thicker than the
focal plane[349]. Figure A5.
Appendices Experimental Results and Discussion
191
Figure A5, Principle of confocal microscopy
• Cytology: the study of cells as described in e Medicine Dictionary. It is branch of
life science, which deals with the study of cells in terms of structure, function and
chemistry.
• Cytopathology: the study of cellular disease and the use of cellular changes for the
diagnosis of disease.
• Cell biology: the study of (normal) cellular anatomy, function and chemistry.
• Densitometry: is the quantitative measurement of optical density in light-sensitive
materials, such as photographic paper or photographic film, due to exposure to
light.
• Ductal carcinoma: is the most common type of breast cancer in women. It comes in
two forms: invasive ductal carcinoma (IDC), ductal carcinoma in situ (DCIS),
(lactiferous ducts), where breast cancer most often originates.
• Dysplasia: is an ambiguous term used in pathology to refer to an abnormality of
development or an epithelial anomaly of growth and differentiation, as explained at
Dorland's Medical Dictionary.
• The endothelium: This is the thin layer of cells that lines the interior surface of
blood vessels and lymphatic vessels as explained at Dorland's Medical Dictionary.
• Fourier transform spectroscopy: is a measurement technique whereby spectra are
collected based on measurements of the coherence of a radiative source, using
time-domain or space-domain measurements of the electromagnetic radiation or
other type of radiation[350].
• Glycogen: is made and stored primarily in the cells of the liver and the muscles. It
functions as the secondary long-term energy storage (with the primary energy
stores being fats held in adipose tissue)[351].
Appendices Experimental Results and Discussion
192
• Histology: It is commonly performed by examining cells and tissues by sectioning
and staining which is followed by examination under a light microscope or electron
microscope.
• Histopathology: in clinical medicine, histopathology refers to the examination of a
biopsy or surgical specimen by a pathologist, after the specimen has been processed
and histological sections have been placed onto glass slides.
• In situ: is a Latin phrase that translates literally to 'in position’, as described in
Collins Latin Dictionary & Grammar, it is used in many different contexts.
• in vivo: (Latin) are those in which the effects of various biological entities are
tested on whole, living organisms usually animals including humans, and plants as
opposed to a partial or dead organism.
• Karyometry: the measurement of the nucleus of a cell.
• Malignant: Many cancer treatments work by coaxing malignant cells to commit
suicide, a process known to biologists as apoptosis.
• Metastasis: is the spread of a cancer from one organ or part to another non-adjacent
organ or part.
• Microarray: A microarray is a multiplex lab-on-a-chip. It is a 2D array on a solid
substrate (usually a glass slide or silicon thin-film cell) that assays large amounts of
biological material using high-throughput screening miniaturized, multiplexed and
parallel processing and detection methods.
• Nucleolus: is a structure found in the nucleus of cells. It forms around specific
chromosomal regions in the nucleus of eukaryotic cells, and is made up of proteins
and ribonucleic acids. Figure A6.
Figure A6 showed the nucleolus is contained within the cell nucleus. (Source: Wikimedia Foundation, Inc.)
Appendices Experimental Results and Discussion
193
• Optical Density (OD): is a convenient tool to describe the transmission of light
through a highly blocking optical filter (when the transmission is extremely small).
OD is defined as the negative of the logarithm (base 10) of the transmission, where
the transmission varies between 0 and 1. Figure A7
OD = – log10 (T) or T = 10 – OD
• Pathology: is a significant component of the causal study of disease and a major field in
modern medical practice and diagnosis. It is a study of disease in general, "general
pathology," it includes a number of distinct but inter-related medical specialties which
diagnose disease mostly through the analysis of tissue, cell, and body fluid samples.
Figure A8.
Figure A7 displayed the negative of the logarithm (base 10) of the optical Density (OD). (Source: [email protected])
Figure A8, showing a pathologist examines a tissue section for evidence of cancerous cells while a surgeon observes. (Source: wikipedia.org)
• Premalignancy: is a generalized state associated with a significantly increased risk
of cancer. If it is left untreated, these conditions may lead to cancer.
• Stroma: the connective, functionally supportive framework of a biological cell,
tissue, or organ.
Appendices Experimental Results and Discussion
194
• Transmission electron microscopy (TEM): is a microscopy technique in which a
beam of electrons is transmitted through an ultra-thin specimen, An image is
formed from the interaction of the electrons transmitted through the specimen
which is magnified and focused onto an imaging device, such as a fluorescent
screen, on a layer of photographic film, or to be detected by a sensor such as a CCD
camera. Figure A9.
Optical density (OD) provides a linear relationship between image intensity and staining
density which can be used for transmission microscopy of stained tissue. In order to get
this, firstly, the intensity of light transmitted through a specimen is calculated using
Lambert- Beer's law for the atmosphere is usually written in equation B1[352]:
* B.1
Where is the intensity of the observed light, the intensity of incident light, is the
amount of stain, is the absorption factor of the stain, and is the distance travelled
through the sample [353] [354, 355]. After this, reduces this relationship to a linear one
equation B2,[353] [354] [355]:
B.2
Is calculated on a channel-by-channel basis. There are many different applications of
OD as well other than stained medical imagery. In general, the equations are generally
specified in a similar fashion whenever the absorption of light by a material is of
importance.
Figure A9, showed: a) A TEM image of the polio virus is 30 nm in size, b) Layout of Optical components in a basic TEM. (Source: Wikimedia)Appendix B: Optical Density of
Transmission Microscopy Images
For instance, image and the corresponding image can be seen in Figure B1 (a).
The incident light intensity for each channel is estimated by the brightest pixel. It is
Appendices Experimental Results and Discussion
195
noteworthy that, in the image, brighter pixels correspond (linearly) to larger amounts
of stain. There are two stains present in the image presented in Figure B1, i.e. hematoxylin
and eosin. Each stain has a different absorption factor , left over from the original
formulation of Beer- Lambert's law. One cannot directly separate the individual
contributions of each stain without images of singly-stained tissue, but will instead get an
estimate of the total stain present.
However, the analysis performed in[354], presents relative contributions of hematoxylin
and eosin to the and channels as [0.1 0.2 0.08] and [0.01 0.13 0.01], respectively.
It can be observed that these relative contributions are qualitatively expressed in the optical
density image of Figure B1. The individual channels of Figure B1 (b) are shown in B1(c)-
(e) as reference. The hematoxylin (which stains nuclei) is represented in both the red and
green channels, eosin (which stains cytoplasm and fibres) is represented mostly in the
green channel, and the blue channel has very little response for either of the stain; this
corroborates with the quantitative responses presented in[353].
Appendices Experimental Results and Discussion
196
B. Appendix B. Glossary of Pathology Terminology
While the optical density corresponding to a 3-channel RGB image (Figure B 1) is being
presented here, the same computation can also be performed for each channel of a
multispectral image stack. This operation is widely used in the histopathology analysis.
Figure B1. Example optical density image.
Appendices Experimental Results and Discussion
197
C. Appendix C: Glossary of Machine Learning and Computer Vision Terminology
• Accuracy: In the fields of science, engineering, industry, and statistics, the
accuracy of a measurement system is the degree of closeness of measurements of a
quantity to the actual (true) value of that quantity[356]
• Classification: Classification can be referred to categorization. It is the process in
which ideas and objects are recognized, differentiated, and understood.
• In graph theory: a connected component or component: is a sub graph
(undirected graph ) in which any two vertices are connected to each other by paths,
and which is connected to no additional vertices in the subparagraph, as displayed
in figure C1.
Figure C1, showed a graph with three connected components
• Cycle: is a simplifiedsimplicial chain with 0 boundaries.
• Densitometry: is the quantitative measurement of optical density due to exposure
to light in light-sensitive materials, such as photographic paper or photographic
film.
• Detection rate: is considered the most important characteristic of a Satellite-AIS
system.
• Eigen decomposition: In the mathematical discipline of linear algebra, Eigen-
decomposition is the factorization of a matrix into a canonical form, whereby the
matrix is represented in terms of its eigenvalues & eigenvectors. Only
diagonalizable matrices can be factorized in this way [357].
Appendices Experimental Results and Discussion
198
• Entropy: Its originates from machine engineering, where in a process involving
heat it is a measure of the portion of heat becoming unavailable for doing work, e.g.
reversal of the process just mentioned. Entropy is also frequently used with
emphasis of order or disorder properties of dissimilar areas.
• Error: It is a measure of the estimated difference between the observed or
calculated value of a quantity and its respective true value.
• False alarm rate: In statistics, performing multiple comparisons, the term false
positive ratio, also known as the false alarm ratio, usually refers to the probability
of falsely rejecting the null hypothesis for a particular test.
• Feature selection: In machine learning figure C2 and statistics, feature selection,
known as variable selection, attribute selection or variable subset selection, is the
process of selecting a subset of relevant features for use in the model construction.
A feature selection technique is based on the central assumption that the data
contains many redundant or irrelevant features.
Figure C2, displayed the machine learning and data mining which used to solve problem in the areas of Classification, Clustering, Regression, Anomaly detection, Association
rules, Reinforcement learning, Structured prediction, Feature learning, Online learning, Semi-supervised learning, and Grammar induction.
• Fractal dimension: A fractal dimension is a ratio providing a statistical index of
complexity comparing how much detail in a pattern (fractal pattern) changes with
the scale at which it is measured. It has also been characterized as a measure of the
space-filling capacity of a pattern that tells how a fractal scales differently from the
space it is embedded in; a fractal dimension does not have to be an integer[358]
[359] [360].
• Gabor filter: In image processing, a Gabor filter, named after Dennis Gabor, is a
linear filter that is used for edge detection. Frequency and orientation
representations of Gabor filters are similar to those of the human visual system.
Appendices Experimental Results and Discussion
199
They have been found to be particularly appropriate for texture representation and
discrimination. In the spatial domain, a 2D Gabor filter is a Gaussian kernel
function modulated by a sinusoidal plane wave.
• Gabor wavelet: Gabor filters are directly related to Gabor wavelets, since they can
be designed for a number of dilations and rotations. However, in general, expansion
is not applied for Gabor wavelets, because this requires computation of bi-
orthogonal wavelets, which may be very time-consuming. Therefore, usually, a
filter bank consisting of Gabor filters with various scales and rotations is created.
The filters are convolved with the signal, resulting in a so-called Gabor space. This
process is closely related to processes in the primary visual cortex[361]. The Gabor
space is very useful in image processing applications such as optical character
recognition, iris recognition and fingerprint recognition. Relations between
activations for a specific spatial location are very distinctive between objects in an
image. Furthermore, important activations can be extracted from the Gabor space in
order to create a sparse object representation. This is an example implementation in
MATLAB/Octave:
Function gb =gabor_fn (sigma, theta, lambda, psi, gamma)
sigma_x = sigma; sigma_y = sigma/gamma;
• Graph: is a representation of a set of objects where some pairs of objects are
connected by links. The interconnected objects are represented by mathematical
abstractions called vertices, and the links that connect some pairs of vertices are
called edges[362]. Typically, a graph is depicted in diagrammatic form as a set of
dots for the vertices, joined by lines or curves for the edges. Graphs are one of the
objects of study in discrete mathematics.
• Ground truth: In machine learning, the term "ground truth" refers to the accuracy
of the training set's classification for supervised learning techniques. This is used in
statistical models to prove or disprove research hypotheses. The term "ground
trothing" refers to the process of gathering the proper objective data for this test.
• In-sample: makes it easy to expect medical costs and health outcomes in your
patient population.
Appendices Experimental Results and Discussion
200
• Optical density: is a logarithmic ratio of the falling radiation to the transmitted
radiation through a material. It is also referred as a fraction of absorbed radiation at
a particular wavelength.
• Out-of-sample: is a popular way to test the expected accuracy of an estimating
method. It is balanced since it’s unprotected of all adjustments and filters.
• K-means: is a method of vector quantization, originally from signal processing,
that is popular for cluster analysis in data mining. K-means clustering aims to
partition n observations into k clusters in which each observation belongs to the
cluster with the nearest mean, serving as a prototype of the cluster. This results in a
partitioning of the data space into Voronoi cells.
• Laplacian of Gaussian: is a 2-D isotropic measure of the 2nd spatial derivative of
an image. The Laplacian of an image highlights regions of rapid intensity change
and is therefore often used for edge detection. The Laplacian is often applied to an
image that has first been smoothed with something approximating a Gaussian
smoothing filter in order to reduce its sensitivity to noise, and hence the two
variants will be described together here. The operator normally takes a single
graylevel image as input and produces another graylevel image as output.
• Margin: in machine learning, it is the distance between a decision boundary and a
data point.
• Mark-up: A (document) mark-up language is a modern system for annotating a
document in a way that is syntactically distinguishable from the text.
• Over segmentation techniques: For some images it is not possible to set
segmentation process parameters, such as a threshold value, so that all the objects
of interest are extracted from the background or each other without over
segmenting the data. Over segmentation is the process by which the objects being
segmented from the background are they segmented or fractured into
subcomponents.
• Pattern spectrum: the application of pattern spectra algorithms to images of
materials of different types, for the purpose of pattern classification. As materials
are often best characterized by their texture, pattern spectra constitute a very
important tool for texture analysis. Researchers by experiments discussed pattern
spectra and they compare the algorithms in terms of speed and accuracy. From the
Appendices Experimental Results and Discussion
201
results it is evident that the slope pattern spectra algorithm is a fast and robust
alternative to granulometric-based techniques.
• Principal Components Analysis (PCA): is a statistical procedure that uses an
orthogonal transformation to convert a set of observations of possibly correlated
variables into a set of values of linearly uncorrelated variables called principal
components. The number of principal components is less than or equal to the
number of original variables. This transformation is defined in such a way that the
first principal component has the largest possible variance (that is, accounts for as
much of the variability in the data as possible), and each succeeding component in
turn has the highest variance possible under the constraint that it is orthogonal to
(i.e., uncorrelated with) the preceding components. Principal components are
guaranteed to be independent if the data set is jointly normally distributed. PCA is
sensitive to the relative scaling of the original variables[363].
• Receiver Operating Characteristic (ROC) curve: In signal detection theory, a
receiver operating characteristic (ROC), or simply ROC curve is a graphical plot
which illustrates the performance of a binary classifier system as its discrimination
threshold is varied. It is created by plotting the fraction of true positives out of the
total actual positives (TPR = true positive rate) vs. the fraction of false positives out
of the total actual negatives (FPR = false positive rate), at various threshold
settings. TPR is also known as sensitivity or recall in machine learning. The FPR is
also known as the fall-out and can be calculated as one minus the more well-known
specificity. The ROC curve was first developed by electrical engineers and radar
engineers during World War II for detecting enemy objects in battlefields and was
soon introduced to psychology to account for perceptual detection of stimuli. ROC
analysis since then has been used in medicine, radiology, biometrics, and other
areas for many decades and is increasingly used in machine learning and data
mining research [364].
• Segmentation: Image segmentation; mean the partitioning of a digital image into
two or more regions.
• Under segmentation: A outcome case to the over segmentation problem [365] is
that of under segmentation. In under segmentation, segmenting processes are
applied to extract objects of interest from an image as overlapping objects, rather
than separate components. The best solution to the under segmentation problem is
Appendices Experimental Results and Discussion
202
to avoid it. For example, the techniques of region growing[365] can be used as a
segmentation method to prevent under segmentation. However, even these methods
might produce under segmented results with some classes of data.
• Watershed transform: used different approaches for image segmentation, such as:
Local minima of the gradient of the image may be chosen as markers, in this case
an over-segmentation is produced and a second step involves region merging. And
marker based watershed transformation make use of specific marker positions
which have been either explicitly defined by the user or determined automatically
with morphological operators or other ways[366].
Appendices Experimental Results and Discussion
203
D. Appendix D: Margin between two hyper planes
Figure D1 suppose the margin is m, it can be computed as equation D.1:
D.1
This is the projection of in the direction of . The equation above can be
rewritten as equation D.2:
D.2
Considering Equations. 3.2 and Equation. D.2 can be written as in equation D.3 & D.4:
D.3
= 2/|w| D.4
Figure D1: Margin m between two supporting hyperplanes
Appendices Experimental Results and Discussion
204
E. Appendix E: Lagrangian dual optimization
Considering Eq. 3.13, the construction of the Lagrangian for the optimization problem is
displayed in equation E.1 & E.2:
E.1
+ E.2
This gives the Lagrangian optimization problem as displayed on equation E.3:
E.3
Subjects to:
E.4
Since the objective function is rounded and the restraint is linear, the solution for and
has to satisfy the following KKT conditions, equations E.5 – E.9:
Gradient condition E.5
E.6
Orthogonally condition
Feasibility condition
Non-negative condition E.9
The Lagrangian optimization can be solved through the partial derivative as displayed in
the following equations E.10 – E.12:
E.10
E.11
E.12
The Lagrangian dual optimization problem, displayed as in equation E.13:
E.13
Bias can be determined as in equation E.14:
E.14
Appendices Experimental Results and Discussion
205
Finally, the optimal decision surface can be defined as displayed in equation E.15 and
E.16:
E.15
E.16
Appendices Experimental Results and Discussion
206
F. Appendix F: Soft-margin nonlinear support vector machine
= F.1
The Lagrangian (is named after Italian-French mathematician, it’s a function that descripts
the state of a dynamic system in terms of position coordinates and their time derivatives
and that is equal to the…) structure of the optimization problem is modified as displayed in
equation F.2:
- F.2
This gives the Lagrangian optimization problem as displayed in equation F.3:
F.3 Subject to the limits:
F.4
F.5 Since the objective function is rounded and the limit is linear, the solution for ,
has to satisfy the first order necessary conditions for a solution in mathematics, the Karush-
Kuhn-Tucker (KKT) conditions (also known as the Kuhn-Tucker) conditions as displayed
in equations F.6 – F.14:
L F.6
L/ L F.7
L/ L F.8
F.9
F.10
F.11
F.12
0 F.13
Appendices Experimental Results and Discussion
207
F.14 The partial derivatives are displayed in equations F.15 - F.19:
F.15
F.16
= = F.17
= C - - F.18
F .19 Substituting partial derivatives above to the Lagrangian of the optimization problem gives
the Lagrangian dual as displayed in equation F.20:
F.20
This Lagrangian dual for soft-margin SVM is same with the Lagrangian dual for hard-
margin SVM. Therefore the objective functions of the optimization problems for both
SVMs are same. The difference is in the limitation. In the soft-margin SVM, the constraint
can be stated as in equation F.21:
F.21
Appendices Experimental Results and Discussion
208
G. Appendix G: Sequential minimal optimization (SMO) for SVM
The sequential minimal optimization (SMO) is used to find the Lagrangian multiplier in
Equation 3.30, which is the solution displayed on equation G.1:
G.1
Where
G.2
subjects to:
G.3 And
= G.4 This maximization problem is equivalent to a minimization problem by multiplying it
by , so that it becomes the minimization problem as displayed in equations G.5 & G.6:
G.5
G.6 Where
G.7 Or
G.8
Thus, to find , minimization of in Eq. G.8 is used. Can be written in another form
as displayed in equation G.9:
G.9
where:
- = is the vector of all ones.
Appendices Experimental Results and Discussion
209
- =
- is a symmetric matrix ( is the number of training data, with components:
- G.10
Furthermore, the constraint defined in Equation 3.31, can be written in the
other form as displayed in equation G.11:
G.11
Where is the vector with components , It’s reported that the difficulty in the minimization of is that Q is a thick matrix
which crops a problem of memory and is time consuming[367]. SMO solves this problem
by decomposition such that at each step two Lagrangian multipliers are optimized and
update the SVM at the end of each optimization. Researchers in[212] described the
algorithm as follows:
a) Defining for the initial solution (iteration n = 1).
b) If is a stationary point of the minimization of , stop. Otherwise, a two–
element working set is determined.
c) Defining (the set contains all those elements of which are
not in ).
d) Let and be the sub-vectors of corresponding to and , respectively.
e) Considering to Equation G.9, solve for the following sub-problem with the
variable :
Where
.12
=
.13
=
G.14
= + G.15
Appendices Experimental Results and Discussion
210
= [ ] + [ ] + G.16
f) Set to be the optimal solution of the sub-problem Equation G.16 and
.Set and go to Step b.
The index is updated at every iteration, but for simplicity, is used instead of .
The stopping criteria consider that the feasible solution of is a stationary point of
Equation G.9 (or in another form, Equation G.10 if and only if there is a number and
two non-negative and such that reported in [212].
G.17
With the limitations: = G.18 Where is the gradient of . Equations G.17 and G.18 result displayed in equation G.19:
G.19
Because of that considering in Equation G.19, can be expressed as in equation G.20:
G.20 Where
And G.21
And,
, , G.22
- Therefore, a feasible is a stationary point of problem in Equation G.10 if and only if:
.23 and the stopping principles can be defined as in equation G.24: G.24 where is tolerance; typically ,
Appendices Experimental Results and Discussion
211
H. Appendix H: Glossary and Abbreviations
ABCD-E: Asymmetry, Border, Colour, Diameter, Evaluation
AC : Accuracy
ACO : Ant Colony Optimization
ADC : Analog to digital Convertor.
ADF : Anisotropic Diffusion Filters
AFE : Automated Feature Extraction
ALOS : Average Length Of Stay
AMF : Adaptive Median Filter
AMPCo : Australian Medical Publishing Company
ANN : Artificial Neural Network
ANS : Autonomic Nervous System
ART : Adaptive Resonance Theory
BCC : Basal Cell Carcinoma.
BI : Binary Images.
B: M : Benign to malignant ratio
BNN : Backpropagation neural network
CA : Concavity Alignment
CAD : Computer-aided design.
CDF : Cumulative Distribution Function.
CBIR : Content–Based Image Retrieval System.
CCD : Charge Coupled Device
CLSM : Confocal laser scanning microscopy
CMOS : Complementary Metal-Oxide Semiconductor.
CNS : Central nervous system
CSE : Clinical skin examination
CSIRO : Commonwealth Scientific and Industrial Research Organization
Appendices Experimental Results and Discussion
212
CT : CAT scan
DALY : Disability-Adjusted life year
DCIS : Ductal Carcinoma In Situ
DIA : Digital Image Acquisition
DIR : Digital Image Representation
DNA : Deoxyribonucleic Acid.
DOG : Difference Of Gaussian
DR : Detection Rate
DWT : Discrete Wavelet Transform
2D-DWT : Two-Dimensional Discrete Wavelet Transform
EEG : Electroencephalogram Test to Measure Brain Electrical: Activity
ELM : Epiluminescence Light Microscopy
EM : Electron microscope
EM : Expectation Maximization
EP : Extra Pixels
FAR : False Alarm Rate
FCM : Fuzzy C-means
FFT : Fast Fourier transforms.
FIM : Feature Intensity Map
FIS : Fuzzy Inference System
FL : Fuzzy Logic
FN : False Negative
FNAC : Fine-Needle Aspiration Cytology
FP : False Positive
FPR : False Positive Rate
FS : Feature Selection
GA : Genetic Algorithm
GF : Gabor Filter.
GFBA : Gabor Filter Bank Application.
GHE : Gray-scale Histogram Equalization
GM : Geometric Mean
GP : General Practitioner Physicians
GT : Ground Truth
Appendices Experimental Results and Discussion
213
GVF : Gradient Vector Flow (Snakes)
H.E. : Histogram Equalization.
H&E : Haematoxylin & Eosin
HIS : Hybrid Intelligent System
HPF : High Pass Filter
HP : Histogram Processing.
HSI : Hue, Saturation, Intensity
HSL : Hue-Saturation-Lightness.
HSV : Hue, Saturation, Value
IDC : Invasive Ductal Carcinoma
IDS : International Dermoscopy Society
IHP : Independent Histogram Pursuit
II : Intensity Images.
IPT : Image Processing Toolbox.
JSEG : Japan Society of Engineering geology.
KKT : Karush-Kuhn-Tucker
K-NN : k-Nearest Neighbour
LAHE : Local Area Histogram Equalization.
LHE : Local Histogram Equalisation
LM : Light Microscope
LMS : Least Mean Square
LoG : Laplacian of Gaussian
LMS : Least Mean Square
LOO : Leave-one-out
LSVM : Linear Support Vector Machine
MACWM: Modified Adaptive Centre Weighed Median
MACWM : Modified Adaptive Canter Weighted Median
MDS : Multi-Dimensional Scaling
MED : Minimum Euclidean Distance
MF : Median Filter.
.MI : Medical Imaging
MIR : Mortality-to-Incidence Ratio
MM : Malignant Melanoma.
Appendices Experimental Results and Discussion
214
MOD : Mean Optical Density
MRI : Magnetic Resonance Imaging
MRF : Markov Random Field
MRMR : Minimal-Redundancy-Maximal-Relevance
NHMRC : National Health and Medical Research Council
NLSVM : Nonlinear Support Vector Machine
NMSC : Non-Melanoma Skin Cancer.
NN : Neural Networks
OD : Optical Density
OM : Optical Microscope
OS-FCM : Orientation Sensitive Fuzzy C-Means.
ORSCSS : Oxford Recognition Skin Cancer Screen System
ORL : Oxford Recognition Limited
P : Performance
PCA : Principal Components Analysis
PDF : Probability Density Function.
PNS : Peripheral nervous system
PSO : Particle Swarm Optimization
RACGP : Royal Australian College of General Practionrs
RBF : Radial Basis Function
RCM : Reflectance Confocal Microscopy.
RFE : Recursive Feature Elimination
RGB : Red, Green, Blue
RGBI : Red, Green and Blue images
RL : Reinforcement Learning
ROC : Receiver Operating Characteristics
egion Of Interest
RTD : Relative tumour density
SAQ : Self-administered questionnaire
SBS : Sequential Backward Selection
SCC : Squamous Cell Carcinoma
SE : Sensitivity
SEM : Scanning Electron Microscope
Appendices Experimental Results and Discussion
215
SF : Spatial Filtering.
SP : Specificity
SFS : Sequential Forward Selection
SIT : Skin Imaging Techniques
SMO : Sequential Minimal Optimization
SMR : Swarm Based Multiple Regression
SN : Spinal nerves
SR : Segmentation Regions
SRM : Statistical Region Merging
SSE : Skin self-examination
warm Based Support Vector Machine
SVM : Support Vector Machine
SSVM : Swarm Optimization Support Vector Machine
TDV : Total Dermoscopy Value
TEM : Transmission Electron Microscope
TLM : Transillumination Light Microscopy.
TN : True Negative
TP : True Positive
TPR : True Positive Rate
US : United States of America
UV : Ultraviolet Radiation
UVR : Ultraviolet radiation
VGP : Vertical Growth Phase
WF : Wiener Filter.
WLG Wavelet Gabor
WPT : Wavelet Packet Transform
YLD : Years of healthy life lost due to disability
YLL : Years of Life Lost
Appendices Experimental Results and Discussion
216
Glossary [77] Aboriginal or Torres Strait Islander: A person of Aboriginal and/or Torres Strait
Islander descent that identifies as an Aboriginal and/or Torres Strait Islander. See also
Indigenous.
Additional diagnosis: A condition or complaint either coexisting with the principal
diagnosis or arising during the episode of care.
Administrative databases: Observations about events that are routinely recorded or
required by law to be recorded. Such events include births, deaths, hospital separations and
cancer incidence. Administrative databases include the Australian Cancer Database, the
National Mortality Database and the National Hospital Morbidity Database.
Admitted patient: A person who undergoes a hospital’s formal admission process to
receive treatment and/or care. Such treatment or care can occur in hospital and/or in the
person’s home (as a ‘hospital-in-home’ patient).
Age-specific rate: A rate for a specific age group. The numerator and denominator relate
to the same age group.
Age-standardisation: A method of removing the influence of age when comparing
populations with different age structures. This is usually necessary because the rates of
many
diseases vary strongly (usually increasing) with age. The age structures of the different
populations are converted to the same ‘standard’ structure; then the disease rates that
would have occurred with that structure are calculated and compared.
Average length of stay (ALOS): The average (mean) number of patient days for admitted
patient episodes. Patients admitted and separated on the same date are allocated a length of
stay of 1 day.
Benign: Non-cancerous tumours that may grow larger but do not spread to other parts of
the body.
Burden of disease and injury: Term referring to the quantified impact of a disease or
injury on an individual or population, using the disability-adjusted life year (DALY)
measure.
Cancer (malignant neoplasm): A large range of diseases in which some of the body’s
cells become defective begin to multiply out of control, can invade and damage the area
around them, and can also spread to other parts of the body to cause further damage.
Appendices Experimental Results and Discussion
217
Chemotherapy: The use of drugs (chemicals) to prevent or treat disease, with the term
being applied for treatment of cancer rather than for other uses.
Cohort method: A method of calculating survival that is based on a cohort of people
diagnosed with cancer in a previous time period and followed over time.
Disability-adjusted life years (DALYs): A year of healthy life lost, either through
premature death or equivalently through living with disability due to illness or injury. It is
the basis unit used in burden of disease and injury estimates.
Death due to cancer: A death where the underlying cause is indicated as cancer.
Expected survival: A measure of survival that reflects the proportion of people in the
general population alive for a given amount of time. Expected survival estimates are crude
estimates calculated from life tables of the general population by age, sex and calendar
year.
Histology: The microscopic characteristics of cellular structure and composition of tissue.
Incidence: The number of new cases (of an illness or event, and so on) in a given period.
Indigenous: A person of Aboriginal and/or Torres Strait Islander descent who identifies as
an Aboriginal and/or Torres Strait Islander. See also Aboriginal or Torres Strait Islander.
Length of stay: Duration of hospital stay, calculated by subtracting the date the patient
was admitted from the day of separation. All leave days, including the day the patient went
on leave, are excluded. A same-day patient is allocated a length of stay of 1 day.
Life tables: Tables of annual probabilities of death in the general population.
Limited-duration prevalence: The number of people alive at a specific time who have
been diagnosed with cancer over a specified period (such as the previous 5 or 25 years).
Malignant: A tumour with the capacity to spread to surrounding tissue or to other sites in
the body. 186 Cancer in Australia: an overview, 2012
Median: The midpoint of a list of observations that have been ranked from the smallest to
the largest.
.Mortality due to cancer: The number of deaths which occurred during a specified period
(usually a year) for which the underlying cause of death was recorded as cancer.
Mortality-to-incidence ratio (MIR): The ratio of the age-standardised mortality rate for
cancer to the age-standardised incidence rate for cancer.
Neoplasm: An abnormal (‘neo’, new) growth of tissue. Can be ‘benign’ (not a cancer) or
‘malignant’ (a cancer). Also known as a tumour.
Appendices Experimental Results and Discussion
218
Non-Indigenous: People who have declared that they are not of Aboriginal or Torres
Strait Islander descent.
Observed survival: A measure of survival that reflects the proportion of people alive for a
given amount of time after a diagnosis of cancer. Observed survival estimates are crude
estimates calculated from population-based cancer data.
Other Australians: Includes people who have declared that they are not of Aboriginal or
Torres Strait Islander descent as well as those who have not stated their Indigenous status.
Overnight patient: An admitted patient who receives hospital treatment for a minimum of
1 night (that is, is admitted to, and separates from, hospital on different dates).
Patient days: The number of full or partial days of stay for patients who were admitted for
an episode of care and who underwent separation during the reporting period. A patient
who is admitted and separated on the same day is allocated one patient day.
Period method: A method of calculating survival that is based on the survival experience
during a recent at-risk or follow-up time period.
Primary cancer: A tumour that is at the site where it first formed (see also Secondary
cancer).
Principal diagnosis: The diagnosis listed in hospital records to describe the problem that
was chiefly responsible for the patient’s episode of care in hospital.
Procedure: A clinical intervention that is surgical in nature, carries a procedural risk,
carries an anaesthetic risk, requires specialised training and/or requires special facilities or
equipment available only in the acute care setting.
Relative survival: The ratio of observed survival of a group of persons diagnosed with
cancer to expected survival of those in the corresponding general population after a
specified interval following diagnosis (such as 5 or 10 years).
Risk factor: Any factor that represents a greater risk of a health disorder or other
unwanted condition or event. Some risk factors are regarded as causes of disease, others
are not Cancer in Australia: an overview, 2012 187 necessarily so. Along with their
opposites, namely protective factors, risk factors are known as ‘determinants’.
Same-day patient: A patient who is admitted to, and separates from, hospital on the same
date.
Secondary cancer: A tumour that originated from a cancer elsewhere in the body. Also
referred to as a metastasis.
Separation: An episode of care for an admitted patient which may include a total hospital
Appendices
219
stay (from admission to discharge, transfer or death) or a portion of a hospital stay that
begins or ends in a change of type of care (for example, from acute to rehabilitation). In
this report, separations are also referred to as hospitalisations.
Statistical significance: An indication from a statistical test that an observed difference or
association may be significant or ‘real’ because it is unlikely to be due just to chance. A
statistical result is usually said to be ‘significant’ if it would occur by chance only once in
20 times or less often (see Appendix B for more information about statistical significance).
Stage: The extent of a cancer in the body. Staging is usually based on the size of the
tumour, whether lymph nodes contain cancer, and whether the cancer has spread from the
original site to other parts of the body.
Survival: A general term indicating the probability of being alive for a given amount time
after a particular event, such as a diagnosis of cancer.
Symptom: Any indication of a disorder that is apparent to the person affected.
Underlying cause of death: The disease or injury that initiated the sequence of events
leading directly to death.
Valid FOBT test: Only faecal occult blood test (FOBT) results that are either positive or
negative are classified as valid results. Inconclusive results are excluded from analysis.
Years of healthy life lost due to disability (YLD): For each new case of cancer, YLD
equals the average duration of the cancer (to remission or death) multiplied by a severity
weight for cancer (which depends upon its disabling effect over the disease duration).
Years of life lost (YLL): For each new case, YLL equals the number of years between
premature death and the standard life expectancy for the individual.
Bibliography
220
Bibliography
1. Curado, M.P., et al., Cancer incidence in five continents. 2007, Intl Agency for Research on Cancer: Lyon,. p. 1-895.
2. Government, A. Cancer in Australia statistics. 2014; 1-3]. Available from: (http://canceraustralia.gov.au).
3. “Cancer Facts & Figures 2002”, in American Cancer Society. 2002: Atlanta, Georgia.
4. Aziz, N. and R.J. . Trends and advances in cancer survivorship research: Challenge and opportunity. in Seminars in Radiation Oncology. 2003.
5. Aziz, N., Cancer survivorship research: Challenge and opportunity. Journal Nutr., 2002. 132(Suppl.11): p. 3494S-503S.
6. Criscione, V. and M. Weinstock, Melanoma Thickness Trends in the United States, 1988-2006. Journal of Investigative Dermatology, 2009.
7. Balch, C.M., S. Soong, and M.I. Ross, Long-term results of a multi-institutional randomized trial comparing prognostic factors and surgical results for intermediate thickness melanomas (1.0 to 4.0 mm). Intergroup Melanoma Surgical Trial. Ann Surg Oncol, 2000. 7: p. 87-97.
8. Barnhill, R.L., J.A. Fine, and G.C. Roush, Predicting fiveyear outcome for patients with cutaneous melanoma in a population-based study. Cancer. Cancer, 1996. 78: p. 427-432.
9. Argenziano G, S.H., Chimenti S, Talamini R, Corona R, Sera F, Dermoscopy of pigmented skin lesions: results of a consensus meeting via the Internet. J Am Acad Dermatol, 2003. 48(5): p. 679-93.
10. Jablonski, N.G., Chaplin, G. , Human skin pigmentation as an adaptation to UV radiation. Proceedings of the National Academy of Sciences, 2010. 107.
11. Stratigos, A.J., et al., Euromelanoma: a dermatology-led European campaign against nonmelanoma skin cancer and cutaneous melanoma. Past, present and future. British Journal of Dermatology, 2012. 167: p. 99-104.
12. Blackledge, J.M. and D.A. Dubovitskiy, Object Detection and Classification with Applications to Skin Cancer Screening. ISAST TRANSACTIONS ON INTELLIGENT SYSTEMS, 2008. 1(2): p. 1-12.
13. Blum, A., Jaworski,S., Clear differences in hand-held dermoscopes. JDDG, 2006. 4: p. 1054–1057.
14. Boucheron, L.E., Object- and Spatial-Level Quantitative Analysis of Multispectral Histopathology Images for Detection and Characterization of Cancer, in Electrical
Bibliography
221
and Computer Engineering. 2008, UNIVERSITY of CALIFORNIA: Santa Barbara. p. 321.
15. Abu Mahmoud, M.K. and A. Al-Jumaily, A Hybrid System for Skin Lesion Detection: Based on Gabor Wavelet and Support Vector Machine, in CISP, BMEI 2014, IEEE, Editor. 2014, IEEE: Dalian, China.
16. (NRHA), N.R.H.A. Skin Cancer in Australia: awareness, early diagnosis and management. National Rural Health Conference Australian Journal of Rural Health 2014; Available from: http://ruralhealth.org.au/sites/default/files/documents/nrha-policy-document/submissions/sub-skin-cancer-inquiry-21-march-2014.pdf.
17. Rigel DS, C.J., Malignant melanoma: prevention early detection, and treatment in the 21st century. CA Cancer J Clin, 2000. 50: p. 215-36.
18. Authors, A., Cancer Facts & Figures 2002. American Cancer Society, 2002: p. 1-44.
19. Chiem, A., A. Al-Jumaily, and R.N. Khushaba. A Novel Hybrid System for Skin Lesion Detection. 2007. IEEE.
20. Abu Mahmoud, M.K., Al-Jumaily,A., Segmentation of Skin Cancer Images Based on Gradient Vector Flow (GVF) Snake, in 2011 IEEE International Conference on Mechatronics and Automation ICMA. 2011. p. 216-220.
21. Abu Mahmoud, M.K., Al-Jumaily, A., Takruri, M., Wavelet and Curvelet Analysis for Automatic Identification of Melanoma Based on Neural Network Classification, in International Journal of Computer Information Systems and Industrial Management (IJCISIM). 2013, IJCISIM: MIR Labs, USA. p. 606-614.
22. Abu Mahmoud, M.K., et al., Classification of Malignant Melanoma and Benign Nevi from Skin Lesions based on Support Vector Machine, in Computational Intelligence, Modelling and Simulation (CIMSim), 2013 Fifth International Conference. 2013: Seoul, Korea (South). p. 236 - 241.
23. Abu Mahmoud, M.K. and A. Al-Jumaily, Novel feature extraction methodology based on histopathalogical images and subsequent classification by Support Vector Machine., in IEEE, I.C. Publications, Editor. 2014. p. 1 - 6
24. Takruri, M., A. Al-Jumaily, and M.K. Abu Mahmoud, Automatic Recognition of Melanoma Using Support Vector Machines: A Study Based on Wavelet, Curvelet and Color Features, in IEEE. 2014.
25. Abu Mahmoud, M.K. and A. Al-Jumaily, A Hybrid System for Skin Lesion Detection: Based on Gabor Wavelet and Support Vector Machine, in IEEE. 2014, IEEE.
26. Jamarani, S.M.H., G. Rezai-rad, and H. Behnam, A novel method for breast cancer prognosis using wavelet packet based neural network, in 27th Annual International Conference of the Engineering in Medicine and Biology Society 2005, (IEEE-EMBS ’05: Shanghai, China.
27. Gonzalez, R.C., R.E. Woods, and S.L. Eddins, Digital Image Processing Using Matlab, ed. 109876543. 2004, Upper Saddle River, New Jersey, USA: Pearson, Prentice Hall.
Bibliography
222
28. Dietterich, T.G., Approximate statistical tests for comparing supervised classification learning algorithms. Neural Computing & Applications, 1997. 10: p. 1895.
29. Abu Mahmoud, M.K., A. Al-Jumaily, and M. Takruri, Wavelet and Curvelet Analysis for Automatic Identification of Melanoma Based on Neural Network Classification, in International Journal of Computer Information Systems and Industrial Management (IJCISIM). 2013, IJCISIM: MIR Labs, USA. p. 606-614.
30. Takruri, M., A. Al-Jumaily, and M.K.A. Mahmoud. Automatic recognition of melanoma using Support Vector Machines: A study based on Wavelet, Curvelet and color features. in Industrial Automation, Information and Communications Technology (IAICT), 2014 International Conference on. 2014. IEEE.
31. Abu Mahmoud, M.K. and Al-Jumaily.A., Novel feature extraction methodology based on histopathalogical images and subsequent classification by Support Vector Machine, in (ICCVIA2014) International Conference on Computer Vision & Image Analysis, Ras Al Khaimah, UAE. 2014.
32. Abu Mahmoud, M.K., et al., Classification of Malignant Melanoma and Benign Nevi from Skin Lesions based on Support Vector Machine, in IEEE 5th International conference on Computational Intelligence, Modelling and Simulation, I.c. socity, Editor. 2013, CIMSim2013: South Korea. p. 236-241.
33. Abu Mahmoud, M.K., A. Al-Jumaily, and M. Takruri, The Automatic Identification of Melanoma by Wavelet and Curvelet Analysis: Study Based on Neural Network Classification, in 11th International Conference on Hybrid Intelligent Systems (HIS). 2011, IEEE. p. 680-685.
34. Abu Mahmoud, M.K. and A. Al-Jumaily, Segmentation of Skin Cancer Images Based on Gradient Vector Flow (GVF) Snake, in 2011 IEEE International Conference on Mechatronics and Automation ICMA. 2011. p. 216-220.
35. Young, B., et al., Wheater's Functional Histology, ed. t.e. 2006. 2006: Elsevier.
36. Ross, M.H., G.I. Kaye, and W. Pawlina, Histology: a text and atlas with cell and molecular biology. 2003.
37. Kelly, J. and D. Van Essen, Cell structure and function in the visual cortex of the cat. The Journal of physiology, 1974. 238(3): p. 515-547.
38. Lodish, H.F., et al., Molecular cell biology. Vol. 4. 2000: WH Freeman New York.
39. Albrecht-Buehler, G., Daughter 3T3 cells. Are they mirror images of each other? The Journal of cell biology, 1977. 72(3): p. 595-603.
40. Young, B., Wheater's Functional Histology: A and Colour Atlas (with student consult online access), . Functional Histology Wheater's, ed. 6e. 2013.
41. Stolz W, B.-F.O., Bilek P, Landthaler M Color alas of drmatoscopy. Cognetta AB. Vol. 2nd ed. 2002, Berlin: Burgdorf WHC.
42. Norton, M.E., N.C. Rose, and P. Benn, Noninvasive prenatal testing for fetal aneuploidy: clinical assessment and a plea for restraint. Obstetrics & Gynecology, 2013. 121(4): p. 847-850.
Bibliography
223
43. L., G. and G.W.G. Kumar Special Stains and H & E Second Edition. Pathology education guide, 2010.
44. TutorVista, Skin Cell Structure. tutoring sessions.
45. Michael, M., Human Evolutionary Biology. Cambridge University Press, 2010: p. 192–213.
46. Lucas, R., McMichael,T., Smith,W.,Armstrong,B., Solar Ultraviolet Radiation: Global burden of disease from solar ultraviolet radiation. Environmental Burden of Disease Series, 2006. 13.
47. Christe, D., Biology. Biology, 2012,2013: p. 1-9.
48. McGrath, J., R. Eady, and F. Pope, Anatomy and organization of human skin.
49. Giovannucci, E.L., Y., et al., Prospective study of predictors of vitamin D status and cancer incidence and mortality in men. JNCI Journal of the National Cancer Institute, 2006. 98(7): p. 451.
50. Staples, M., et al., Non-melanoma skin cancer in Australia. Medical Journal Australia, 2006. 198(184 ).
51. Foundation, S.C. Skin Cancer Foundation. Available from: http://www.skincancer.org.
52. Diepgen, T.L. and V. Mahler, The epidemiology of skin cancer. British Journal of Dermatology, 2002. 146(s61): p. 1-6.
53. Staples, M., R. Marks, and G. Giles, Trends in the incidence of non-melanocytic skin cancer (NMSC) treated in Australia 1985-1995: are primary prevention programs starting to have an effect? International Journal of Cancer, 1998. 78: p. 144-148.
54. Terrill, P., Fairbanks, S., Bailey, M., Is there just one lesion? The need for whole body skin examination in patients presenting with non-melanocytic skin cancer. ANZ Journal of Surgery, 2009. 79(10): p. 707-712.
55. Welfare, A.I.o.H.a., Australian Association of cancer Registratries, in 23, AIHW, Editor. 2003, AIHW: canberra.
56. Randell, P., Cole,J., Quirk, C., Von Nida,J., Yiasemides, E., and Queensland Health The ABCD of melanoma detection. Cancer Council Western Australia, 2009.
57. Morton, D.L., et al., Sentinel node biopsy for early-stage melanoma: accuracy and morbidity in MSLT-I, an international multicenter trial. Annals of surgery, 2005. 242(3): p. 302.
58. Feng, J., et al., Metabolites Studies of Secondary Melanoma on C57BL/6J Mouse Liver Using 1H NMR Metabolomics. 2013. 3(4): p. 1011-1035.
59. Jerant AF, J.J., Sheridan CD, Caffrey TJ "Early detection and treatment of skin cancer". Am Fam Physician, 2000. 62(2): p. 357–68, 375–6, 381–2.
60. Armstrong, B., et al., Challenges in health and health care for Australia. Medical Journal of Australia, 2007. 187(9): p. 485.
Bibliography
224
61. Ferlay, J., Shin HR, Bray, F, Forman, D, Mathers, C, Parkin, DM Cancer Incidence and Mortality Worldwide: IARC Cancerbase No.10. GLOBOCAN 2008 & International Agency for Research on Cancer, 2008. 2.
62. Gies, P., C. Roy, and P. Udelhofen, Solar and ultraviolet radiation. In Prevention of Skin Cancer, 2004: p. 21-54.
63. Hershkovitz L, S.J., Treves AJ, Besser MJ "Focus on adoptive T cell transfer trials in melanoma". Clin. Dev. Immunol. 2010: 260267, 2010.
64. Wurm EM, S.H., Scanning for melanoma. Australian Prescriber 2010. 33: p. 150–5.
65. Hocker, T. and H. Tsao, Ultraviolet radiation and melanoma: a systematic review and analysis of reported sequence variants. Human mutation, 2007. 28(6): p. 578-588.
66. James, W.D.B., Timothy G., Diseases of the Skin: clinical Dermatology., ed. Andrews. 2006: Saunders Elsevier. 694–9.
67. Wisconsin, M.C.o., "Malignant Melanoma: staging". Collaborative Hypertext of Radiology, 2006.
68. Balch C, B.A., Soong S, Atkins M, Cascinelli N, Coit D, Fleming I,, Houghton A, Kirkwood J, McMasters K, Mihm M, Morton D, Reintgen D, Ross M, Sober A, Thompson J, Thompson J, "Final version of the American Joint Committee on Cancer staging system for cutaneous melanoma". Gershenwald J and J Clin Oncol, 2001. 19( 16 ): p. 3635–48.
69. Weedon, D., Skin pathology, 2nd Edition 2002, Sydney: Churchill-Livingstone.
70. Breslow, A., Thickness, cross-sectional areas and depth of invasion in the prognosis of cutaneous melanoma. Ann Surg, 1970. 172: p. 902-908.
71. Aziz, N., Rowland JH. . . Trends and advances in cancer survivorship research: Challenge and opportunity. in Seminars in Radiation Oncology. 2003.
72. Jemal, A., Murray, T., Ward, E, Samuels, A., Tiwari, RC, Ghafoor, A., Feuer, EJ, Thun, MJ, Cancer Statistics 2005. CA Cancer Journal Clin., 2005. 55: p. 10-30.
73. Rowland, J., Mariotto, A.., Aziz, N., Tesauro, G., Feuer E. , Cancer survivorship United States 1971-2001. MMWR, 2004. 53: p. 526-9.
74. Statistics, A.B.o., National Health Survey: Summary of Results 2004-2005 2006. 2008, Australian Bureau of Statistics.
75. Cancer Council Australia. Screening and early detection of skin cancer. 2007; Available from: www.cancer.org.auCancer
76. Cancer, I.A.f.R.o., Cancer incidence in five continents, in International Agency for Research on Cancer. 2002: Lyon.
77. Welfare, A.I.o.H.a., Cancer in Australia: An overview, 2012, in AIHW. 2012: Canberra, Australia.
78. Brady, M.S., Melanoma and Other Skin Cancers. 2013.
79. A., N., et al., Skin cancer-related prevention and screening behaviors: a review of the literature. J Behav Med, 2009. 32: p. 406-428.
Bibliography
225
80. Valiukevi ien , S., Melanom diagnostika ir gydymas. Kaunas, KMU, 2002: p. 85p.
81. Blackledge, J., Fellow, IET, Dubovitskiy, DA, Object Detection and Classification with Applications to Skin Cancer Screening. Intelligent Systems, 2008. 1(1): p. 34-45.
82. Wyszecki, G., Stiles, W.S., Color Science, Second Edition. McGraw-Hill, NY. 1982.
83. Stockham, T.G., Image Processing in the context of a Visual Model. Jr. Proceedings of IEEE, 1972. 60(7): p. 828-842.
84. Demirkaya, O., M.H. Asyali, and P.K. Sahoo, Image Processing with Matlab: Applications in medicine and biology. 2009: CRC Press.
85. Pham, D.T., Alcock, R., Smart Inspection Systems: Technicuqes and Applications of Intelligent Vision. Academic Press, Oxford. 2003.
86. Lissesand, T.M., Kiefer, R.W. , Remote Sensing and Image Interpretation. 1999: 4th Edition, John Wiley and Sons.
87. Jensen, J.R., Remote Sensing of the Environment: An Earth ResourcePerspective. 2000: Prentice Hall.
88. Suetens, P., Fundamentals of Medical Imaging. Cambridge University Press. 2002.
89. Kupinski, A.A., Giger, h4.. Automated Seeded Lesion Segmentation on Digital Mammograms. IEEE Trans. Med. Image, 1998. 17: p. 510-51 7.
90. Mitra, S.S. and T. Acharya, Data Mining: Multimedia. Soft Computing, and Bioinformatics. Vol. NJ. 2003: Wiley, Hoboken.
91. Acharya, T. and A.K. Ray, Information Technology: Principles and Applications. prentice Hall of India. 2004, New Delhi, India.
92. Cucchiara, R., et al., The Sakbot System for Moving Object Detection and Tracking. Video-Based Surveillance Systems- Computer Vision and Distributed Processing, 2001: p. 145-157.
93. Davies, E.R., Machine Vision: Theory, Algorithms, Practicalities. 1997, Academic press: London.
94. Freeman, J.H., Machine vision. Algorithms, Architectures, and Systems, in Academic press. 1988, Academic press: London,.
95. Louis, J., Galbiati, JR, Machine vision and digital image processing fundamentals. State University of New York, New-York, 1990.
96. Snyder, W.E. and H. Qi., Machine Vision. 2004, Cambridge University Press: England.
97. Rosenfeld, A. and A.C. Kak, Digital Picture Processing. Second Edition, Academic Press. Vol. 1. 1982.
98. Bracewell, R.N., Two-Dimensional Imaging. Prentice Hall, Englewood Cliffs, NJ, 1995.
99. Toennies, K.D., Guide to Medical Image Analysis: Methods and Algorithms. 2012, London: Springer-Verlag.
Bibliography
226
100. Hall, C.F., Hall, E.L., A Nonlinear Model for the Spatial Characteristics of the Human Visual System. IEEE Trans. Systems, Man, and Cybernetics, SMC, 1977. 7(3): p. 161-170.
101. Bayer, B.E., Color Imaging Array. Eastman Kodak Company, US Patent, 1976. 3: p. 971,065.
102. Sakamoto, T., C. Nakanishi, and T. Hase, Software Pixel Interpolation for Digital Still Cameras Suitable for A 32-bit MCU. IEEE Transactions on Consumer Electronics, 1998. 44(4): p. 1342-1352.
103. wilhelm , P., Enzo,B., Peter,E., Maibach,I., , “Bio Engineering of Skin: Skin Imaging and Analysis”, Dermatology: Basic science series. Dermatology: Basic science series., 2008.
104. Marozas, M., Review on skin lesion imaging, analysis and automatic classification. R.J. Biomedical Engineering Institute, 2006.
105. DS., R., Malignant melanoma: perspectives on incidence and its effects on awareness, diagnosis, and treatment. CA Cancer J Clin, 1996. 46: p. 195-8.
106. Australia, A.G.C., Melanoma of the skin statistics. Information Publication Scheme, 2013.
107. Bafounta ML, B.A., Aegerter P, Saiag P, Is dermoscopy (epiluminescence microscopy) useful for the diagnosis of melanoma? Results of a meta-analysis using techniques adapted to the evaluation of diagnostic tests. Arch Dermatol, 2001. 137
p. 1343-50.
108. Rigel DS, F.R., Kopf AW, The incidence of malignant melanoma in the United States: Issues as we approach the 21st entury. J Am Acad Dermatol, 1996. 34: p. 839-47.
109. Stolz W, B.-F.O., Bilek P, Landthaler M, Blackwell Wissenschafts-Verlag; Cognetta AB. Color alas of drmatoscopy. 2nd ed. 2002, Berlin: Burgdorf WHC.
110. Binder, M., Puespoeck-Schwarz, Steiner, A., Kittler, H., Muellner, M., Wolff, A.,Pehamberger, H. , Austria, A., Epiluminescence Microscopy. “A useful tool for the diagnosis of pigmented lesions for formally trained dermatologists”. Journal of the American Academy of Dermatology, 1997. 36(2): p. 197-202.
111. Binder M, S.M., Winkler A, Steiner A, Kaider A, Wolff K,, Epiluminescence microscopy. A useful tool for the diagnosis of pigmented skin lesions for formally trained dermatologists. Arch Dermatol, 1995. 131: p. 286-91.
112. Kittler H, P.H., Wolff K, Binder M, Diagnostic accuracy of dermoscopy. Lancet Oncol, 2002. 3: p. 159-65.
113. Pehamberger, H., Steiner,A., Wolff,K., In vivo epiluminescence microscopy of pigmented skin lesions. i. pattern analysis of pigmented skin lesions. Journal of the American Academy of Dermatology, 1987. 17(4): p. 571–583.
114. Tennessee, U.o., Comparison of OM, TEM and SEM. Electron Microscopy, ed. D.o.M.S.a.E. (ed.). Vol. Topics 3b,c. University of Tennessee.
Bibliography
227
115. Fink-Puches, R.H.-W.R., J. Smolle, and H. Kerl, Confocal laser scanning microscopy: a new optical microscopic technique for applications in pathology and dermatology. Cutan Pathol, 1995. 22(3): p. 252-9.
116. Chu, M., N. Kollias, and M.A. Luedtke, Confocal Scanning Laser Microscopy: Applications for Imaging Dynamic Processes Skin In Vivo, 2010.
117. A., S., et al., Statistical evaluation of epiluminescence microscopy criteria for melanocytic pigmented skin lesions. J. Am. Acad. Derm., 1993. 29(4): p. 581-588.
118. Steiner A, P.H., Wolff K, In vivo epiluminescence microscopy of pigmented skin lesions. II. Diagnosis of small pigmented skin lesions and early detection of malignant melanoma. J Am Acad Dermatol, 1987 17 p. 584-91.
119. Nachbar F, S.W., Merkle T, Cognetta AB, Vogt T, Landthaler M, The ABCD rule of dermatoscopy: high prospective value in the diagnosis of doubtful melanocytic skin lesions. J Am Acad Dermatol, 1994. 30: p. 551-9.
120. Robert, C. and M.D. Burton, Malignant Melanoma in the Year 2000. CA Cancer J Clin 2000, 2000. 50(4): p. 209-213.
121. Menzies, S.W.C.K. and I. C., An atlas of surface microscopy of pigmented skin lesions, ed. M.W. 2003. 2003, Sydney: McGraw-Hill Book Company.
122. Argenziano, G., et al., Dermoscopy of pigmented skin lesions: results of a consensus meeting via the Internet. J Am Acad Dermatol, 2003. 48(5): p. 679-93.
123. Menzies SW, I.C., McCarthy WH. , A sensitivity and specificity analysis of the surface microscopy features of invasive melanoma. Melanoma Res 1996 6: p. 55-62.
124. Braun RP, R.H., Oliviero M, Kopf AW, Saurat JH., Pattern analysis: a two step procedure for the dermoscopic diagnosis of melanoma. Clin Dermatol, 2002. 20: p. 236-9.
125. Simon, J.D., et al., Current challenges in understanding melanogenesis: bridging chemistry, biological control, morphology, and function. Pigment cell & melanoma research, 2009. 22(5): p. 563-579.
126. Hansen, C., Wilkinson, D., Hansen, M., Argenziano, G., How good are skin cancer clinics at melanoma detection?Number needed to treat variability across a national clinic group in Australia. American Acadimic Dermatol 2009(Am Acad Dermatol,).
127. Kreusch J, R.G., Standardisierte auflichtmikroskopische Unterscheidung melanozytischer und nichtmelanozytischer Pigmentmale. Hautarzt, 1991(42): p. 77-83.
128. Stolz W, R.A., Cognetta AB, ABCD rule of dermatoscopy: a new practical method for early recognition of malignant melanoma. Eur J Dermatol, 1994: p. 4:521-527.
129. Staples, M.P., Elwood, M., Burton, RC., Nonmelanoma skin cancer in Australia: the 2002 national survey and trends since 1985. Medical Journal Austustralia, 2006. 184: p. 6-10.
130. Anthony, F. and M.D. Jerant, Early detection and Treatment of Skin cancer. Am Fam Physician, 2000. 62(2): p. 357-368.
Bibliography
228
131. MenziesS.W., et al., The performance of solarscan: an automated dermoscopy image analysis instrument for the diagnosis of primary melanoma. Archives of dermatology, 2005. 141(11): p. 1388–1396.
132. Zalaudek, I., et al., Three point checklist of dermoscopy: an open internet study. British journal of dermatology, 2006. 154(3): p. 431-437.
133. Celebi, M.E., Development of Algorithms for Dermoscopy Image Analysis. 2006: ProQuest.
134. Sydney, B.D., Skin lession boundary tracing algorithm. Melanoma-skin cancer reviewed, 2006.
135. Celebi, M.E., et al., Unsupervised border detection in dermoscopy images. Skin Research and Technology and Blackwell Munksgaard, 2007. 13: p. 454-462.
136. Zouridakis, G., M. Doshi, and N. Mullani, Early diagnosis of skin cancer based on segmentation and measurement of vascularization and pigmentation in nevoscope images, in 26th Annual International Conference of the IEEE EMBS. 2004.
137. Winter, A. and C. Nastar, Differential feature distribution maps for image segmentation and region queries image databases, 1999: p. 1-5.
138. Ruzon, M.A., Segmentation. 1997.
139. Ma, W. and B. Manjunath, Texture features and learning similarity. Computer Vision and Pattern Recognition, 1996. 5: p. 425-430.
140. Zarita, R. and S. Lelandais, Wavelets and high order statistics for texture classification. SCIA, 1997. 8: p. 95-102.
141. Jain, A.K., Murty, M. N., Flynn, P. J. , Data clustering: a review, in American Computer Magazin Computing Surveys (CSUR). 1999. p. 1-60.
142. Squire, D., Data mining - clustering techniques and association rule discovery. 2005.
143. Chaira, T. and A.K. Ray, Fuzzy Image Processing and Applications with MatLab. 2010, U.S.: Taylor & Francis Group.
144. Wolfram, fuzzy clustering. 2005. 1 p. 12
145. Schmid, P., Segmentation and symmetry measure for image analysis: application to digital dermatoscopy., in PRÉSENTÉE AU DÉPARTEMENT D'ÉLECTRICITÉ. 1999, ÉCOLE POLYTECHNIQUE FÉDÉRALE DE LAUSANNE: Lausanne, EPFL. p. 192.
146. Erkol, B., et al., Automatic lesion boundary detection in dermoscopy images using gradient vector flow snakes. Skin Research and Technology, 2004. 11: p. 17-26.
147. Geman, S. and D. Geman, Stochastic relaxation and gibbs distributions and the bayesian restoration of images. IEEE-PAMI, 1984. 6: p. 721-741.
148. Deng, H. and D.A. Clausi, Unsupervised image segmentation using a simple mrf model with a new implementation scheme. Pattern Recognition, 2004. 37: p. 2323-2335.
Bibliography
229
149. Crowson, A.N., C.M. Magro, and M.C. Mihm, Prognosticators of melanoma, the melanoma report, and the sentinel lymph node. Modern Pathology, 2006. 19: p. S71-S87.
150. Clark Jr, W.H., A classification of malignant melanoma in man correlated with histogenesis and biologic behavior. Advances in the Biology of the Skin, 1967. 8: p. 621-647.
151. Clark Jr WH, From L, and B. EA., Histogenesis and biologic behavior of primary human malignant melanomas of the skin. Cancer Res, 1969. 29: p. 705-726.
152. Crowson, A., et al., The pathology of melanoma. Cutaneous Melanoma, 4th edn. Quality Medical Publishers Inc.: St Louis, 2003: p. 171-206.
153. Starfield, B., et al., The effects of specialist supply on populations’ health: assessing the evidence. Cancer, 2005. 103(20.93): p. 23.18.
154. Babu, M.N., et al., Histo-pathological Image Analysis using OS-FCM and Level Sets. 2010.
155. Demir, C. and B. Yener, Automated cancer diagnosis based on histopathological images: a systematic survey, in TR-05-09. 2009, Rensselaer Polytechnic Institute, Department of computer science.
156. Mulrane, L., et al., Automated image analysis in histopathology: a valuable tool in medical diagnostics. 2008.
157. Clark Jr WH, D.E. Elder, and D. Guerry, Model predicting survival in stage I melanoma based on tumor progression. Natl Cancer Inst, 1989. 81: p. 1893-1904.
158. Clark, W.H. and M.C. Mihm, Lentigo maligna and lentigo maligna melanoma. Am J Pathol, 1969. 55: p. 39-54.
159. Reed, R.J., Acral lentiginous melanoma. In: New Concepts in Surgical Pathology of the Skin. Wiley: New York, 1976: p. 89-90.
160. Koh, H.K., E. Michalik, and A.J. Sober, Lentigo maligna melanoma has no better prognosis than other types of melanoma.blogsome.com. J Clin Oncol 1984. 2: p. 994-1001.
161. Vollmer, R.T., Malignant melanoma. A multivariate analysis of prognostic factors. Pathol Annu, 1989. 24: p. 383-407.
162. Weinstock, M.A. and A.J. Sober, The risk of progression of lentigo maligna to lentigo maligna melanoma. Br J Dermatol, 1987. 116: p. 303-310.
163. Fidler, I.J., Tumor heterogeneity and the biology of cancer invasion and metastasis. Cancer Res, 1978. 38: p. 2651-2660.
164. Balch, C.M., S.J. Soong, and J.E. Gershenwald, Prognostic factors analysis of 17,600 melanoma patients: validation of the American Joint Committee on Cancer melanoma staging system. J Clin Oncol, 2001. 19: p. 3622-3634.
165. Balch, C.M., T.M. Murad, and S.J. Soong, A multlifactorial analysis of melanoma: prognostic histopathological features comparing Clark’s and Breslow’s staging methods. Ann Surg, 1978. 188: p. 732-742.
Bibliography
230
166. Balch, C.M., S.J. Soong, and T.M. Murad, A multifactorial analysis of melanoma. II. Prognostic factors in patients with stage I (localized) melanoma. Surgery, 1979. 86: p. 343-351.
167. Cascinelli, N., A. Morabito, and R. Bufalino, Prognosis of stage I melanoma of the skin. Int J Cancer, 1980. 26: p. 733-739.
168. Sondergaard, K. and G. Schou, Survival with primary cutaneous malignant melanoma, evaluated from 2012 cases. A multivariate regression analysis. Virchows Arch A Pathol Anat Histopathol, 1985. 406: p. 179-195.
169. Rigel, D.S., R.J. Friedman, and A.W. Kopf, Factors influencing survival in melanoma. . Dermatol Clin, 1991. 9: p. 631-642.
170. Stidham, K.R., J.L. Johnson, and H.F. Seigler, Survival superiority of females with melanoma. A multivariate analysis of 6383 patients exploring the significance of gender in prognostic outcome. Arch Surg, 1994. 129: p. 316-324.
171. Garbe, C., P. Buttner, and J. Bertz, Primary cutaneous melanoma. Identification of prognostic groups and estimation of individual prognosis for 5093 patients. Cancer, 1995. 75: p. 2484-2491.
172. Retsas, S., K. Henry, and M.Q. Mohammed, Prognostic factors of cutaneous melanoma and a new staging system proposed by the American Joint Committee on Cancer (AJCC): validation in a cohort of 1284 patients. Eur J Cancer, 2002. 38: p. 511-516.
173. Ronan, S.G., M.C. Han, and T.K. Das Gupta, Histologic prognostic indicators in cutaneous malignant melanoma. Semin Oncol, 1988. 15: p. 558-565.
174. Vollmer, R.T. and H.F. Seigler, A model for pretest probability of lymph node metastasis from cutaneous melanoma. Am J Clin Pathol, 2000. 114: p. 875-879.
175. DAY, C.L., et al., Malignant Melanoma: Prognostic Significance of "Msicroscopic Satellites" in the Reticular Dermis and Subcutaneous Fat. MICROSCOPIC SATELLITES IN MELANOMA PRIMARIES, 1981. 194(1): p. 108-112.
176. Day, C.L., T.J. Harrist, and F. Gorstein, Malignant melanoma. Prognostic significance of ‘microscopic satellites’ in the reticular dermis and subcutaneous fat. Ann Surg, 1981. 194: p. 108-112.
177. Harrist, T.J., D.S. Rigel, and J.C.L. Day, ‘Microscopic satellites’ are more highly associated with regional lymph nodes metastases than is primary melanoma thickness. . Cancer, 1984. 53: p. 2183-2187.
178. Leon, P., J.M. Daly, and M. Synnestvedt, The prognostic implications of microscopic satellites in patients with clinical stage I melanoma. Arch Surg, 1991. 126: p. 1461-1468.
179. Rao, U.N., J. Ibrahim, and L.E. Flaherty, Implications of microscopic satellites of the primary and extracapsular lymph node spread in patients with high-risk melanoma: pathology corollary of Eastern Cooperative Oncology Group Trial E1690. J Clin Oncol, 2002. 20: p. 2053-2057.
180. Mascaro, J.M., M. Molgo, and T. Castel, Plasma cells within the infiltrate of primary cutaneous malignant melanoma of the skin. A confirmation of its histoprognostic value. Am J Dermatopathol, 1987. 9: p. 497-499.
Bibliography
231
181. Piepkorn, M., Prognostic factors in cutaneous melanoma. Busam KJ (eds). ed. Pathology of Melanocytic Nevi and Malignant Melanoma, ed. n.e.N. York. 2004: Springer, Barnhill RL. 372-394.
182. Benvenuto-Andrade, C.D., S.W.; Agero, A.L.C.;Scope,A.; Rajadhyaksha, M.; Halpern,A.C.; Marghoob,A.A., Differences Between Polarized Light Dermoscopy and Immersion Contact Dermoscopy for the Evaluation of Skin Lesions Arch Dermatol., 2007. 143: p. 329-338.
183. Argenziano, G., et al., Dermoscopy of pigmented skin lesions: results of a consensus meeting via the Internet. Journal of the American Academy of Dermatology, 2003. 48(5): p. 679-693.
184. Kendler, K.S., et al., Specificity of genetic and environmental risk factors for use and abuse/dependence of cannabis, cocaine, hallucinogens, sedatives, stimulants, and opiates in male twins. American Journal of Psychiatry, 2003. 160(4): p. 687-695.
185. Clark, W.H., et al., The histogenesis and biologic behavior of primary human malignant melanomas of the skin. Cancer research, 1969. 29(3): p. 705-727.
186. McGovern, V., Digital diagnosis: new tool for detecting skin cancer. Environ Health Perspect, 2003. 111(14): p. McGovern, V.
187. MenziesS.W, B., H.L., Talbot, A. Gutenev,T.A.,Avramidis M. ,Wong,, Lo,S.K., Mackellar,g., and V. Skladnev, McCarthy, W., The performance of solarscan: an automated dermoscopy image analysis instrument for the diagnosis of primary melanoma. Archives of dermatology, 2005. 141 (11): p. 1388–1396.
188. Elbaum M, K.A., Rabinovitz HS, Langley RG, Kamino H, Mihm MC Jr, Sober AJ, Peck GL, Bogdan A, Gutkowicz-Krusin D, Greenebaum M, Keem S, Oliviero M, Wang S., Automatic differentiation of melanoma from melanocytic nevi with multispectral digital dermoscopy: a feasibility study. J Am Acad Dermatol, 2001. 44(2): p. 207-18.
189. Gutkowicz-Krusin D, E.M., Jacobs A, Keem S, Kopf AW, Kamino H, Wang S, Rubin P, Rabinovitz H, Oliviero M, Precision of automatic measurements of pigmented skin lesion parameters with a MelaFind(TM) multispectral digital dermoscope. Melanoma Res, 2000. 10(6): p. 563-70.
190. Kopf, A.W., ElbaumM.,Provost,N., The use of dermoscopy and digital imaging in the diagnosis of cutaneous malignant melanoma. Skin Research and Technology (impact factor: 1.71). 10/ /j.1600-0846.1997.tb00152.x, 2006. 3(1): p. 1 - 7.
191. Rigel DS, R.J., Friedman R., The evolution of melanoma diagnosis: 25 years beyond the ABCDs. CA Cancer J Clin., 2010. 60(5): p. 301-16.
192. Talbot, H., Bischof,L., An overview of the Polartechnics SolarScan melanoma diagnosis algorithms. 2003.
193. Philippa, H.Y., Peter, D. Baade, Monika, Janda, Christopher, B. Del Mar, et al.,, Diagnosing skin cancer in primary care: how do mainstream general practitioners compare with primary care skin cancer clinic doctors? Medical Journal of Australia, 2007. 187(4): p. 215.
Bibliography
232
194. Steiner, A., H. Pehamberger, and K. Wolff, Improvement of the diagnostic accuracy in pigmented skin lesions by epiluminescent light microscopy. ANTICANCER RESEARCH, 1987. 7(3 Pt B): p. 433-434.
195. Abbasi, N.R., et al., Early diagnosis of cutaneous melanoma: Revisiting the abcd criteria. JAMA: The Journal of the American Medical Association, 2004. 292(22): p. 2771-2776.
196. Robinson Jk, T.R., SKills training to learn discrimination of abcde criteria by those at risk of developing melanoma. Archives of Dermatology, 2006. 142(4): p. 447-452.
197. Johr, R.H., Dermoscopy: alternative melanocytic algorithms—the ABCD rule of dermatoscopy , menzies scoring method, and 7-point checklist. Clinics in dermatology, 2002. 20(3): p. 240-247.
198. Healsmith, M.F., An evaluation of the revised seven-point checklist for the early diagnosis of cutaneous malignant melanoma. British Journal of Dermatology, 1994. 130(1): p. 48-50.
199. Zalaudek, I.A.G., et al., The Dermoscopy Working, Group, Three-point checklist of dermoscopy: an open internet study. British Journal of Dermatology, 2006. 154(3): p. 431-437.
200. Vestergaard, M.E.M. and S. W., Automated Diagnostic Instruments for Cutaneous Melanoma. Seminars in Cutaneous Medicine and Surgery, 2008. 27(1): p. 32-36.
201. Menzies SW, I.C., C. K., and M. W.H., Frequency and morphologic characteristics of invasive melanomas lacking specific surface microscopic features. Archives of Dermatology, 1996. 132: p. 1178-1182.
202. Carli, P., Reliability and inter-observer agreement of dermoscopic diagnosis of melanoma and melanocytic naevi. European Journal of Cancer Prevention, 1998. 7(5): p. 397-402.
203. Stanganelli, I., Intraobserver agreement in interpretation of digital epiluminescence microscopy. Journal of the American Academy of Dermatology, 1995. 33(4): p. 584-589.
204. Cascinelli, N., A possible new tool for clinical diagnosis of melanoma: The computer. Journal of the American Academy of Dermatology, 1987. 16(2, Part 1): p. 361-367.
205. Hall, P.N., E. Claridge, and D.M. Smith, Computer screening for early detection of melanoma—is there a future? . British Journal of Dermatology, 1995. 132(3): p. 325-338.
206. Cristofolini, M., et al., Diagnosis of cutaneous melanoma: accuracy of a computerized image analysis system (Skin View). Skin Research and Technology, 1997. 3(1): p. 23-27.
207. Rajpara, S.M., Systematic review of dermoscopy and digital dermoscopy/ artificial intelligence for the diagnosis of melanoma. British Journal of Dermatology, 2009. 161(3): p. 591-604.
208. Rosado B, M.S., Harbauer A et al., Accuracy of computer diagnosis of melanoma: a quantitative meta-analysis. Archives of Dermatology, 2003. 139: p. 361-367.
Bibliography
233
209. Celebi, M.E., Y.A. Aslandogan, and P.R. Bergstresser, Unsupervised border detection of skin lesion images. ITCC International Conference onInformation Technology: Coding and Computing, 2005, 2005.
210. Sookpotharom, S. Border Detection of Skin Lesion Images Based on Fuzzy C-Means Thresholding. in 3rd International Conference on Genetic and Evolutionary Computing, 2009. WGEC '09. 2009.
211. Silveira, M. and Marques.J.S. Level set segmentation of dermoscopy images. in 5th IEEE International Symposium on Biomedical Imaging: From Nano to Macro, ISBI 2008. 2008.
212. Celebi, M., Aslandogan, YA, Stoecker, WV, Iyatomi, H, Oka, H, Chen, X, Unsupervised border detection in dermoscopy images. Skin Research and Technology, 2007. 13(4): p. 454-462.
213. Maglogiannis, I., et al., Intelligent Segmentation and Classification of Pigmented Skin Lesions in Dermatological Images, in Advances in Artificial Intelligence,, S. Berlin, Editor. 2006: Heidelberg. p. 214-223.
214. Bulent Erkol, R.H.M., et al., Automatic lesion boundary detection in dermoscopy images using gradient vector flow snakes. Skin Research and Technology, 2005. 11: p. 17-26.
215. Celebi, M.E., H.A. Kingravi, and Y.A. Aslandogan, Nonlinear vector filtering for impulsive noise removal from color images. Journal of Electronic Imaging, 2007. 16(3): p. 033008-033008.
216. Celebi, M.E., G.S. Hitoshi Iyatomi, and W.V. Stoecker, Lesion Border Detection in Dermoscopy Images. Comput Med Imaging Graph, 2009. 33(2): p. 148-153.
217. Abbas, Q., Lesion border detection in dermoscopy images using dynamic programming. Skin Research and Technology, 2011. 17(1): p. 91-100.
218. Gomez, D.D., Independent Histogram Pursuit for Segmentation of Skin Lesions. IEEE Transactions on Biomedical Engineering, 2008. 55(1): p. 157-161.
219. Adelmann, H.G., Butterworth equations for homomorphic filtering of images. Computers in Biology and Medicine, 1998. 28(2): p. 169-181.
220. Celebi, M.H.K. and Lee.J., Fast and accurate border detection in dermoscopy images using statistical region merging. Proc. SPIE Medical Imaging., 2007.
221. Celebi, M.E., et al., Border detection in dermoscopy images using statistical region merging. Skin Research and Technology Journal, 2008. 14: p. 1-7.
222. Sweet, M.R., adaptive and recursive median filtering.
223. Butler, J.K., Methods for improved light microscope microtomy. Biotechnic & Histochemistry, 1979. 54(2): p. 53-69.
224. QinzLi, Dark line detection with line width extraction, in 15th IEEE International Conference on Image Processing, 2008. ICIP 2008. 2008.
225. Criminisi, A., P. Perez, and K. Toyama, Region filling and object removal by exemplar-based image inpainting. IEEE Transactions on Image Processing, 2004. 13(9): p. 1200-1212.
226. Filters.
Bibliography
234
227. Press, W.H., Wiener Filtering (and some Wavelets), in Computational Statistics with Application to Bioinformatics, T.U.o.T.a. Austin, Editor. 2008: Texas at Austin. p. 1-26.
228. Lim, J.S., Two-Dimensional Signal and Image Processing, in Two-Dimensional Signal and Image Processing. 1990, Englewood Cliffs, NJ, Prentice Hall. p. p. 548, equations 9.26, 9.27, and 9.29.
229. J.J. Kulikowski, J.J., S. Mar6elja, and P.O. Bishop, Theory of Spatial Position and Spatial Frequency Relations in the Receptive Fields of Simple Ceils in the Visual Cortex. Biol. Cybern, 1982. 43: p. 187-198.
230. Jain, A. and S. Bhattacharjee, Address block location on envelopes using gabor filters. 1992. 25(12).
231. Jain, A., N. Ratha, and S. Lakshmanan, Object detection using gabor filters. Pattern Recognition, 1997. 30: p. 295-309.
232. Utah, E. Gabor Filters. 2002; Available from: www.cs.utah.edu/~arul/report/node13.html.
233. Yazdi, H.S. and F. Homayouni, Impulsive Noise Suppression of Images Using Adaptive Median Filter International Journal of Signal Processing, Image Processing and Pattern Recognition, 2010. 3(3): p. 1-12.
234. Arakawa, K., Median filter based on fuzzy rules and its application to image restoration. Fuzzy Sets and Systems 1996. 77: p. 3-13.
235. Lin, T.-C. and P.-T. Yu, Partition fuzzy median "lter based on fuzzy rules for image restoration. Fuzzy Sets and Systems 2004. 147: p. 75-97.
236. Dhawan, A.P., Medical Image Analysis, ed. I. press. Vol. ch. 8. 2003, United States of America: Wiley-Interscience.
237. Gonzalez, R.C. and R.E. Woods, “Digital Image Processing”, 2nd ed. Upper Saddle River, 2002.
238. Jain, A.K., Fundamentals of Digital Image Processing. Englewood Cliffs, 1989.
239. Russ, J.C., The Image Processing Handbook, 2nd Edition. Boca Raton. 1995: CRCPress.
240. Jain, R., R. Kasturi, and B.G. Schunck, Machine Vision. 1995, New York: McGraw-Hill.
241. Rosenfeld, A. and A.V. Kak, Digital Picture Processing, 2nd ed. Vol. 1 & 2. 1982, Orlando: Academic Press.
242. Czerwinski, R.N., D.L. Jones, and W.D. O'Brien Jr. Ultrasound speckle reduction by directional median filtering. in Image Processing, 1995. Proceedings., International Conference on. 1995. IEEE.
243. Otsu, N., A Tlreshold Selection Method from Gray-Level Histograms. IEEE T. Syst. Man Cy, 1979. 9: p. 62-66.
244. Dawant, B.M., et al., Image segmentation, in Handbook of Medical Imaging,. Vol. 2: Medical Image Processing and Analysis. 2000, Paress, Bellingha, 2000: SPIE.
Bibliography
235
245. Dhawan, A.P. and L. Arata, Segmentation of medical images through competitive learning Comp, Methods and Prog. in Biomed, 1993. 40: p. 203-215.
246. Cline, H.E., et al., Three-dimentional segmentation of MR images of the head using probabilityand connectivity. Comp. Assis. Tomog., 1990. 14: p. 1037-1045.
247. Zavaljevski, A., et al., Multispectral MR brain image classification. Comp. Med. Image. Graphics Image Proc., 2000. 24: p. 87-98.
248. Reibman, A.R. and L. Nolte, Optimal detection and performance of distributed sensor systems. Aerospace and Electronic Systems, IEEE Transactions on, 1987(1): p. 24-30.
249. Hall, L.O., et al., A comparison of neural network and fuzzy clustering techniques in segmentation magnetic resonance images of the brain. IEEE Trans. Neural Networks, 1992. 3: p. 672-682.
250. Loncaric, S., et al., 3-D image analysis of intracerebral brain hemorrhage. Comp. Methods Prog. Biomed, 1995. 46: p. 207-216.
251. Clarke, L., et al., MRI: Stability of three supervised segmentation techniques. Mag. Res. Image, 1993. 11: p. 95-106.
252. Beni Xuanli, G., A validity measure for fuzzy clustring. IEEE Trans Patt. Analy. Machine Intell., 1991. 133(8).
253. Tarca, A.L., et al., Machine learning and its applications to biology. PLoS computational biology, 2007. 3(6): p. e116.
254. Yao, X., Evolving artificial neural networks. Proceedings of the IEEE, 1999. 87(9): p. 1423-1447.
255. Marchiori, S.C., et al., Neural network based on adaptive resonance theory with continuous training for multi-configuration transient stability analysis of electric power systems. Applied Soft Computing, 2011. 11(1): p. 706-715.
256. Fukushima, K., Neocognitron: A hierarchical neural network capable of visual pattern recognition. Neural networks, 1988. 1(2): p. 119-130.
257. Page, I. Intelligent Image Features Extraction in Knowledge Discovery Systems. 2007-2015; Available from: www.imagefeatures.org/research/.
258. Selvarajah, S. and S.R. Kodituwakku, Analysis and Comparison of Texture Features for Content Based Image Retrieval. International Journal of Latest Trends in Computing (E-ISSN: 2045-5364) 108 2011. 2(1): p. 108-113.
259. Manian, V. Image feature extraction and segmentation. Pattern Recognition; Available from: www.ece.uprm.edu/~manian/ImageAnalysis4.
260. Intelligent Image Features Extraction in Knowledge Discovery Systems. Available from: http://www.imagefeatures.org/research/.
261. Haralick, R.M., K. Shanmugan, and I.H. Dinstein, Textural features for image classification. IEEE Trans. Syst., 1973. SMC-3: p. 610–621.
262. Selvarajah, S. and S.R. Kodituwakku, Analysis and Comparison of Texture Features for Content Based Image Retrieval. International Journal of Latest Trends in Computing, 2011. 2(1): p. 108-113.
Bibliography
236
263. Li, B. and M.Q.H. Meng, Texture analysis for ulcer detection in capsule endoscopy images. Image and Vision Computing, 2009. 27(9): p. 1336–1342
264. Chui, C.K., An Introduction to Wavelets, in Academic Press. 1992: San Diego, Calif, USA.
265. Guyon, I. and A. Elisseeff, An introduction to feature extraction, in Feature Extraction. 2006, Springer. p. 1-25.
266. Lyons, R.G., Understanding Digital Siginal Processing. 2010, Upper Saddle River: Prentice Hall. 984.
267. S., J. and Walker, A Primer on Wavelets and Their Scientific Applications. Paperback, 1999.
268. Gonzalez, R.C., image processing. 1992.
269. Soille, morphological image analysis. 2004.
270. Kruskal and Wish, Multidimensional Scaling (MDS). 1978.
271. Liu and H. Motoda, Feature Extraction, Construction and Selection: A Data Mining Perspective. Kluwer Academic, 1998.
272. Abe, S., Support Vector Machines for Pattern Classification. Springer, New York, 2005.
273. Kohavi, R. and G.H. John, Wrappers for feature subset selection. Artif. Intell. Med, 1997. 97: p. 273–324.
274. D. Wettschereck, D., D.W. Aha, and T. Mohri, A review and empirical evaluation of feature weighting methods for a class of lazy learning algorithms. Artif. Intell. Med, 1997. 11: p. 273–314.
275. Bluma, A.L. and P. Langley, Selection of relevant features and examples in machine learning. Amficial Intelligence, 1997. 97: p. 245-271.
276. Kohavi, R., John, G.H., Wrappers for feature subset selection. Artificial Intelligence 97, 1997: p. 273-324.
277. G.H. John, G.H., R. Kohavi, and K. Pfleger. Irrelevant features and the subset selection problem. in Proceedings of the International Conference on Machine Learning. 1994. New Brunswick.
278. Neter, J., M.H. Kutner, and W. Wasserman, Applied Linear Statistical Models. C.J. Nachtsheim, 1996.
279. Ginsberg, M.L., Essentials of Artificial Intelligence. Morgan Kaufmann, Palo Alto, 1993.
280. I.M. Guyon, I.M., et al., Gene selection for cancer classification using support vector machines. Mach. Learn, 2002. 46: p. 389–422.
281. F. Ojeda, F., J.A.K. Suykens, and B. De Moor, Low rank updated LS-SVM classifiers for fast variable selection. Neural Netw., 2008. 21(2–3): p. 437–449.
282. Weston, S., et al. Feature selection for SVMs. in Proceedings of the Conference on Neural Information Processing Systems. 2000. Denver.
Bibliography
237
283. chapelle, O., et al., Choosing Multiple Parameters for Support Vector Machines. Machine Learning, , , , 2002. 46: p. 131–159.
284. Rasmussen, C.E. and C.K.I. Williams, Gaussian Processes for Machine Learning. MIT Press Cambridge, 2006.
285. Cawley, G.C. Leave-one-out cross-validation based model selection criteria for weighted LS-SVMs. in Proceedings of the International Joint Conference. 2006. Neural Networks, Vancouver.
286. Suykens, J.A.K., et al., Least Squares Support Vector Machines. World Scientific, Singapore,, 2002.
287. Duda, R.O., P.E. Hart, and D.G. Stork, Pattern classification. 2012: John Wiley & Sons.
288. Ooi, C.H., M. Chetty, and S.W. Teng, Differential prioritization in feature selection and classifier aggregation for multiclass microarray datasets. Data Mining and Knowledge Discovery, 2007. 14(3): p. 329-366.
289. Rückstieß, T., C. Osendorfer, and P. van der Smagt, Minimizing data consumption with sequential online feature selection. International Journal of Machine Learning and Cybernetics, 2013. 4(3): p. 235-243.
290. Gunn, S.R., Support vector machines for classification and regression. ISIS technical report, 1998. 14.
291. Duin, R.P., Classifiers in almost empty spaces, in In Pattern Recognition, P.t.I. Conference, Editor. 2000, IEEE. p. 1-7.
292. K. Hoffmann, K., et al., Diagnostic and neural analysis of skin cancer (DANAOS). A multi-centre study for collection and computeraided analysis of data from pigmented skin lesions using digital dermoscopy. Br. J. Dermatol., 2003. 149: p. 801–809.
293. BATUWITA, R. and V. PALADE, FSVM-CIL: fuzzy support vector machines for class imbalance learning. IEEE Transactions on Fuzzy Systems, 2010. 18: p. 558-571.
294. Platt, J.C. Fast training of support vector machines using sequential minimal optimization. In Advances in kernel methods 1999 1999, February]; 185-208].
295. KEERTHI, S.S., et al., Improvements to Platt's SMO algorithm for SVM classifier design. Neural Computation, 2001. 13: p. 637-649.
296. FAN, R.E., P.H. CHEN, and C.J. LIN, Working set selection using second order information for training support vector machines. The Journal of Machine Learning Research, 2005. 6: p. 1889-1918.
297. Chiang, M., Convex Optimization and Lagrange Duality. 2007.
298. Smola, A.J., Regression estimation with support vector learning machines. 1996, Technische Universit: M¨unchen.
299. Vapnik, V., et al., Support vector method for function approximation, regression estimation, and signal processing. Advances in Neural Information Processing Systems, 1997. 9: p. 281– 287.
Bibliography
238
300. Osuna, E., et al., An improved training algorithm for support vector machines. J. Principe 1997: p. 276 – 285.
301. DEL VALLE, Y., et al., Particle swarm optimization: basic concepts, variants and applications in power systems. IEEE Transactions on Evolutionary Computation, 2008. 12: p. 171-195.
302. ASTION, M.L., et al., Overtraining in neural networks that interpret clinical data. Clinical Chemistry, 1993. 39: p. 1998-2004.
303. SIDDIQUI, A.B., et al., Block- Based Pixel Level Multi-Focus Image Fusion Using Particle Swarm Optimization. International Journal of Innovative Computing, Information and Control, 2011. 7: p. 3583 - 3596.
304. Cand`es, E.J. and D.L. Donoho, New tight frames of curvelets and optimal representations of objects with piecewise-C 2 singularities. Comm. on Pure and Appl. Math, 2002. 57: p. 219-26.
305. Australian, C.C. Skin Cancer Facts and Figures. 2010; Available from: <http://www.cancer.org.au/cancersmartlifestyle/SunSmart/Skincancerfactsandfigures.htm>.
306. Welfare, A.c.n.H.A.I.o.H.a. Health system expenditures on cancer and other neoplasms in Australia, 2000–01. 2005; Available from: (http://www.aihw.gov.au).
307. Australian Institute of Health and Welfare, C., Cancer in AUSTRALIA: an overview 2010, in AIHW cat no. CAN 56, s.n. 60, Editor. 2010: Canberra.
308. Sikorski, J. Identification of malignant melanoma by wavelet analysis. in Student/Faculty Research Day. 2004. CSIS, Pace University.
309. Zhang, Z., Stoecker,W., Moss, R. , Border detection on digitized skin tumor images. IEEE Trans. Med. Imaging 2000. 19(11): p. 1128-1143.
310. Scholkopf, B., Tsuda, K., Vert, J.P., kernel Methods in computational biology. MIT Press, 2004.
311. Dolianitis, C., et al., Comparative performance of 4 dermoscopic algorithms by nonexperts for the diagnosis of melanocytic lesions. Archives of dermatology, 2005. 141(8): p. 1008-1014.
312. Kass, M., Witkin,A.,Terzopolous, D., Snakes: active contour models. Int. J. Comput, 1987. 1(4): p. 321-331.
313. Xu, C. and J.L. Prince, Snakes, Shapes, and Gradient Vector Flow. IEEE TRANSACTIONS ON IMAGE PROCESSING, 1998. 7(3): p. 359.
314. Tang, J., et al., Surface extraction and thickness measurement of the articular cartilage from MR images using directional gradient vector flow snake. IEEE Trans. Biomed. Eng, 2006. 52(5): p. 896-907.
315. Tang, J., Acton,S. , Vessel boundary tracking for intravital microscopy via multi- scale gradient vector flow snakes. IEEE Trans. Biomed. Eng., 2004. 51(2): p. 316-324.
316. Arakawa, K., Median filter based on fuzzy rules and its application to image restoration. Fuzzy Sets and Systems, 1996. 77: p. 3-13.
Bibliography
239
317. Kass, M., Witkin,A.,Terzopolous, D., Snakes: active contour models. Int. J. Comput, 1988. 1(4): p. 321-331.
318. Caselles, V., et al., A geometric model for active contours. Numerische Mathematik, 1993. 66: p. 1-31.
319. L.D., C., On active contour models and balloons. CVGIP: Image Understanding, 1991. 53(2): p. 211–218.
320. Xu, C. and J.L. Prince, Gradient Vector Flow: A New External Force for Snakes. IEEE Proc. Conf. on Comp. Vis. Patt. Recog. (CVPR'97), 1997: p. 66-71.
321. Tang, J., A multi-directionGVFsnakeforthesegmentationofskincancerimages. Pattern Recognition Lett, 2009. 42: p. 1172 -- 1179.
322. Nixon, M. and A. Aguado, Feature Extraction and Image Processing. 2002, London, UK: Academic Press is an imprint of Elsevier.
323. Levinson, N., The Wiener RMS (root mean square) error criterion in filter design and prediction. 1947.
324. Lau, H.T. and A. Al-Jumaily, Automatically Early Detection of Skin Cancer: Study Based on Nueral Netwok Classification, in International Conference of Soft Computing and Pattern Recognition. 2009.
325. Sumana, I.J., Image Retrieval Using Discrete Curvelet Transform, in Gippsland School of Information Technolog. 2008, Monash University: Australia. p. 84.
326. Takruri, M., A. Al-Jumaily, and M.K. Abu Mahmoud, Automatic Recognition of Melanoma Using Support Vector Machines: A Study Based on Wavelet, Curvelet and Color Features, in 2014 International Conference on Industrial Automation, Information and Communications Technology (IAICT 2014). 2014, IEEE Indonesia Communications Society Chapter and RAS/CSS (Robotics & Automations Society and Control System Society) Joint Chapter: Grand Inna Kuta Bali Indonesia.
327. Robert-Inacio, F., Dinet, E., An adaptive median filter for colour image processing. Proc. 3rd CGIV, 2006: p. 205-210.
328. Dinet, E. and F. Robert-Inacio, Color Median Filtering: a Spatially Adaptive Filter. Proceedings of Image and Vision Computing, 2007: p. 71-76.
329. Teknomo, K., K-Means Clustering Tutorial. 2007. p. What is K-Means Clustering?
Simply speaking it is an algorithm to classify or to group your objects based on attributes/features into K number of group. K is positive integer number. The grouping is done by minimizing the sum of squares of distances between data and the corresponding cluster centroid. Thus, the purpose of K-mean clustering is to classify the data.
330. Van De Wouwer, G., Weyn, B., Scheunders, P., Jacob, W., Van Marck, E., Van Dyck, D.. Wavelets as chromatin texture descriptors for the automated identification of neoplastic nuclei. PubMed - indexed for MEDLINE, 2000. 197(Pt1): p. 25-35.
331. Ali, S., Smith-Miles, K. A., Improved Support Vector Machine Generalization Using Normalized Input Space. © Springer-Verlag Berlin Heidelberg 2006, 2006: p. 362-371.
Bibliography
240
332. Annessi, G., et al., Sensitivity, specificity, and diagnostic accuracy of three dermoscopic algorithmic methods in the diagnosis of doubtful melanocytic lesions: the importance of light brown structureless areas in differentiating atypical melanocytic nevi from thin melanom
J Am Acad Dermatol, 2007. 56: p. 759-767.
333. Henning, J.S., S.W. Dusza, and S.Q. Wang, The CASH (color, architecture, symmetry, and homogeneity) algorithm for dermoscopy. J AmAcad Dermatol, 2007. 56: p. 45-52.
334. Dolianitis, C., et al., Comparative performance of 4 dermoscopic algorithms by nonexperts for the diagnosis of melanocytic lesions. Arch Dermatol, 2005. 141: p. 1008-1014.
335. Menzies SW1, et al., The performance of SolarScan: an automated dermoscopy image analysis instrument for the diagnosis of primary melanoma. Arch Dermatol, 2005. 141(11): p. 1388-96.
336. Burroni, M., P. Sbano, and G. Cevenini, Dysplastic nevus vs. in situ melanoma: digital dermoscopy analysis. J Dermatol, 2005. 152: p. 679-684.
337. Argenziano, G., et al., Dermoscopy of pigmented skin lesions: results of a consensus meeting via the Internet. J Am Acad Dermatol, 2003. 48: p. 679-693.
338. Menzies, S.W., L. Bischof, and H. Talbot, The performance of SolarScan: an automated dermoscopy image analysis instrument for the diagnosis of primary melanoma. Arch Dermatol 2005. 141: p. 1388-1396.
339. Kecman, V., Learning and Soft Computing: Support Vector Machines, Neural Networks, and Fuzzy Logic Models (Complex Adaptive Systems). 2001, Cambridge, MA, 2001: The MIT Press.
340. Wang, L.P. and X.J. Fu, Data Mining with Computational Intelligence. 2006, Berlin: Springer. 288.
341. Sloin, A. and D. Burshtein, Support Vector Machine Training for Improved Hidden Markov Modeling. Signal Processing, IEEE Transactions, 2008. 56(1): p. 172-188.
342. Chu, F. and L.P. Wang, Applications of support vector machines to cancer classification with microarray data. International Journal of Neural Systems, 2005. 15(6): p. 475-484.
343. del Valle, Y., et al., Particle Swarm Optimization: Basic Concepts, Variants and Applications in Power Systems. IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, 2008. 12(2): p. 171-195.
344. Hsu, C.-W., C.-C. Chang, and C.-J. Lin, A practical guide to support vector classification. 2003.
345. Wallace, V., et al., Classification of reflectance spectra from pigmented skin lesions, a comparison of multivariate discriminant analysis and artificial neural networks. Physics in medicine and biology, 2000. 45(10): p. 2859.
346. GEORGOULAS, G. and C.D. STYLIOS, Predicting the risk of metabolic acidosis for newborns based on fetal heart rate signal classification using support vector machines. IEEE Transactions on Biomedical Engineering, 2006. 53: p. 875-884.
Bibliography
241
347. DEEB, O. and M. GOODARZI, Exploring QSARs for Inhibitory Activity of Non-peptide HIV-1 Protease Inhibitors by GA-PLS and GA-SVM. Chemical Biology and Drug Design, 2010. 75: p. 506-514.
348. Lin, S.-W., et al., Particle swarm optimization for parameter determination and feature selection of support vector machines. Expert Systems with Applications, 2008. 35(4): p. 1817-1824.
349. Pawley, J., Handbook of biological confocal microscopy. 2010: Springer.
350. Infrared spectroscopy from Wikipedia, t.f.e.; Available from: http://en.wikipedia.org/wiki/Infrared_spectroscopy.
351. Tyne, U.o.N.u. On-line medical dictionary. Available from: http://cancerweb.ncl.ac.uk/omd.
352. Mehta, A., Limitations and Deviations of Beer–Lambert Law.
353. Brey, E.M., et al., Automated Selection of DAB-labeled Tissue for Immunohistochemical Quantification. The Journal of Histochemistry & Cytochemistry, 2003. 51(5): p. 575–584.
354. Ruifrok, A.C., Johnston, D.A., Quantification of histochemical staining by color deconvolution. Analytical and quantitative cytology and histology, 2001. 23(4): p. 291-299.
355. Beer-Lambert. law from Wikipedia. Available from: http://en.wikipedia.org/wiki/Beer-Lambert_law.
356. Bipm, I., et al., The international vocabulary of metrology—basic and general concepts and associated terms (VIM). JCGM, 2008. 200.
357. Meyer, C.D., Matrix Analysis and Applied Linear Algebra. SIAM, 2000.
358. Falconer and Kenneth, Fractal Geometry. 2003, New York: Wiley.
359. Sagan and Hans, Space-Filling Curves. 1994, Berlin: Springer-Verlag.
360. Vicsek and Tamás, Fractal growth phenomena. 1992, Singapore New Jersey: World Scientific.
361. Daugman, J.G., Two-dimensional spectral analysis of cortical receptive field profiles. Vision Res, 1980. 20(10): p. 847–56.
362. Trudeau and Richard, Introduction to Graph Theory (Corrected, enlarged republication. ed.), in New York: Dover Pub. 1993, Retrieved 8 August 2012. p. 19.
363. Jolliffe, I.T., Principal Component Analysis, second edition. 2002: Springer.
364. Swet and A. John, Signal detection theory and ROC analysis in psychology and diagnostics : collected papers. NJ, 1996.
365. Vision Systems Design. 1999 4(2): p. 21.
366. Beucher, S. and Meyer.F., The morphological approach to segmentation: the watershed transformation. In Mathematical Morphology in Image Processing (Ed. E. R. Dougherty). 1993. 433–481.
367. Platt, J.C., Fast Training of Support Vector Machines using Sequential Minimal Optimization. 2000, Redmond, WA 98052, USA: Microsoft Way.
Bibliography
242